def createWorld(username='******',
level=0,
ability='good',
explanation='none',
embodiment='robot',
acknowledgment='no',
sequence=False,
root='.',
ext='xml',
beliefs=True):
"""
Creates the initial PsychSim scenario and saves it
@param username: name of user ID to use in filenames
@param level: robot mission level to use as template
@param ability: the level of the robot's ability
- good or C{True}: perfect sensors and sensor model
- badSensor or C{False}: noisy sensors, but perfect model of noisy sensors
- badModel: perfect sensors, but imperfect model of sensors
@type ability: bool
@param explanation: the type of explanation to use
- none: No explanations
- ability: Explanation based on robot ability provided.
- abilitybenevolence: Explanation based on both robot's ability and benevolence provided.
@type explanation: str
@param embodiment: the robot's embodiment
- robot: The robot looks like a robot
- dog: The robot looks like a dog
@type embodiment: str
@param acknowledgment: the robot's behavior regarding the acknowledgment of errors
- no: The robot does not acknowledge its errors
- yes: The robot acknowledges its errors
@type acknowledgment: str
@param root: the root directory to use for files (default is current working directory)
@param ext: the file extension for the PsychSim scenario file
- xml: Save as uncompressed XML
- psy: Save as bzipped XML
@type ext: str
@param beliefs: if C{True}, store robot's uncertain beliefs in scenario file, rather than compute them on the fly. Storing in scenario file makes the scenario a more complete model, but greatly increases file sixe. Default is C{True}
@type beliefs: bool
"""
print "**************************createWorld***********************"
print 'Username:\t%s\nLevel:\t\t%s' % (username, level + 1)
print 'Ability\t\t%s\nExplanation:\t%s\nEmbodiment:\t%s\nAcknowledge:\t%s' % \
(ability,explanation,embodiment,acknowledgment)
# Pre-compute symbols for this level's waypoints
for point in WAYPOINTS[level]:
if not point.has_key('symbol'):
point['symbol'] = point['name'].replace(' ', '')
world = World()
world.defineState(
None,
'level',
int,
lo=0,
hi=len(WAYPOINTS) - 1,
description='Static variable indicating what mission level')
world.setState(None, 'level', level)
world.defineState(None, 'time', float)
world.setState(None, 'time', 0.)
world.defineState(None, 'complete', bool)
world.setState(None, 'complete', False)
world.addTermination(
makeTree({
'if': trueRow('complete'),
True: True,
False: False
}))
# Buildings
threats = ['none', 'NBC', 'armed']
for waypoint in WAYPOINTS[level]:
if not waypoint.has_key('symbol'):
waypoint['symbol'] = waypoint['name'].replace(' ', '')
world.addAgent(Agent(waypoint['symbol']))
# Have we visited this waypoint?
key = world.defineState(waypoint['symbol'], 'visited', bool)
world.setFeature(key, False)
# Are there dangerous chemicals or armed people here?
key = world.defineState(waypoint['symbol'], 'danger', list, threats[:])
if waypoint.has_key('NBC') and waypoint['NBC']:
world.setFeature(key, 'NBC')
elif waypoint.has_key('armed') and waypoint['armed']:
world.setFeature(key, 'armed')
else:
world.setFeature(key, 'none')
key = world.defineState(waypoint['symbol'], 'recommendation', list,
['none', 'protected', 'unprotected'])
world.setFeature(key, 'none')
# Human
human = Agent('human')
world.addAgent(human)
world.defineState(human.name, 'alive', bool)
human.setState('alive', True)
world.defineState(human.name, 'deaths', int)
human.setState('deaths', 0)
# Robot
robot = Agent('robot')
world.addAgent(robot)
# Robot states
world.defineState(robot.name, 'waypoint', list,
[point['symbol'] for point in WAYPOINTS[level]])
robot.setState('waypoint', WAYPOINTS[level][getStart(level)]['symbol'])
world.defineState(robot.name, 'explanation', list, [
'none', 'ability', 'abilitybenevolence', 'abilityconfidence',
'confidence'
])
robot.setState('explanation', explanation)
world.defineState(robot.name, 'embodiment', list, ['robot', 'dog'])
robot.setState('embodiment', embodiment)
world.defineState(robot.name, 'acknowledgment', list, ['no', 'yes'])
robot.setState('acknowledgment', acknowledgment)
world.defineState(robot.name, 'ability', list,
['badSensor', 'badModel', 'good'])
if ability is True:
# Backward compatibility with boolean ability
ability = 'good'
elif ability is False:
ability = 'badSensor'
robot.setState('ability', ability)
# State of the robot's sensors
world.defineState(robot.name, 'sensorModel', list, ['good', 'bad'])
robot.setState('sensorModel', 'good')
world.defineState(robot.name, 'command', list, ['none'] +
[point['symbol'] for point in WAYPOINTS[level]])
robot.setState('command', 'none')
# Actions
for end in range(len(WAYPOINTS[level])):
symbol = WAYPOINTS[level][end]['symbol']
# Robot movement
action = robot.addAction({'verb': 'moveto', 'object': symbol})
# Legal if no contradictory command
tree = makeTree({
'if': equalRow(stateKey(robot.name, 'command'), 'none'),
True: True,
False: {
'if': equalRow(stateKey(robot.name, 'command'), symbol),
True: True,
False: False
}
})
robot.setLegal(action, tree)
# Dynamics of robot's location
tree = makeTree(
setToConstantMatrix(stateKey(action['subject'], 'waypoint'),
symbol))
world.setDynamics(stateKey(action['subject'], 'waypoint'), action,
tree)
# Dynamics of visited flag
key = stateKey(symbol, 'visited')
tree = makeTree(setTrueMatrix(key))
world.setDynamics(key, action, tree)
# Dynamics of time
key = stateKey(None, 'time')
tree = setToConstantMatrix(key, 0.)
for start in range(len(WAYPOINTS[level])):
if start != end:
startsymbol = WAYPOINTS[level][start]['symbol']
if sequence:
# Distance is measured by level sequence
distance = abs(start - end) * 50
else:
try:
distance = DISTANCES[WAYPOINTS[level][start]['name']][
WAYPOINTS[level][end]['name']]
except KeyError:
try:
distance = DISTANCES[WAYPOINTS[level][end][
'name']][WAYPOINTS[level][start]['name']]
except KeyError:
distance = 250
tree = {
'if':
equalRow(stateKey(action['subject'], 'waypoint'),
startsymbol),
True:
setToConstantMatrix(key,
float(distance) / 1000.),
False:
tree
}
world.setDynamics(key, action, makeTree(tree))
# Human entry: Dead or alive if unprotected?
key = stateKey(human.name, 'alive')
action = robot.addAction({
'verb': 'recommend unprotected',
'object': symbol
})
tree = makeTree({
'if': equalRow(stateKey(symbol, 'danger'), 'none'),
True: setTrueMatrix(key),
False: setFalseMatrix(key)
})
world.setDynamics(key, action, tree)
robot.setLegal(action, makeTree(False))
# Human entry: How much "time" if protected?
action = robot.addAction({
'verb': 'recommend protected',
'object': symbol
})
key = stateKey(None, 'time')
world.setDynamics(key, action, makeTree(setToConstantMatrix(key,
0.25)))
robot.setLegal(action, makeTree(False))
# Robot goals
goal = minimizeFeature(stateKey(None, 'time'))
robot.setReward(goal, 2.)
goal = maximizeFeature(stateKey(human.name, 'alive'))
robot.setReward(goal, 1.)
for point in WAYPOINTS[level]:
robot.setReward(maximizeFeature(stateKey(point['symbol'], 'visited')),
1.)
if beliefs:
# omega = 'danger'
world.defineVariable(robot.name, ActionSet)
# Robot beliefs
world.setModel(robot.name, True)
value = 1. / float(len(WAYPOINTS[level]))
# tree = KeyedVector({CONSTANT: world.value2float(omega,'none')})
for index in range(len(WAYPOINTS[level])):
waypoint = WAYPOINTS[level][index]
key = stateKey(waypoint['symbol'], 'danger')
# if index > 0:
# Starting state is safe
robot.setBelief(
key,
psychsim.probability.Distribution({
'NBC': value / 2.,
'armed': value / 2.,
'none': 1. - value
}))
# Observation function
# tree = {'if': equalRow(stateKey(robot.name,'waypoint'),waypoint['symbol']),
# True: generateO(world,key),
# False: tree}
# robot.defineObservation(omega,makeTree(tree),domain=list,lo=['none','NBC','armed'])
robot.defineObservation('microphone',
makeTree(None),
None,
domain=list,
lo=['nobody', 'friendly', 'suspicious'])
robot.defineObservation('NBCsensor', makeTree(None), None, domain=bool)
robot.defineObservation('camera', makeTree(None), None, domain=bool)
else:
robot.defineObservation('microphone',
makeTree(None),
None,
domain=list,
lo=['nobody', 'friendly', 'suspicious'])
robot.defineObservation('NBCsensor', makeTree(None), None, domain=bool)
robot.defineObservation('camera', makeTree(None), None, domain=bool)
robot.setAttribute('horizon', 1)
world.setOrder([robot.name])
filename = getFilename(username, level, ext, root)
world.save(filename, ext == 'psy')
WriteLogData('%s user %s, level %d, ability %s, explanation %s' % \
(CREATE_TAG,username,level,ability,explanation),username,level,root=root)
return world
class TestAgents(unittest.TestCase):
def setUp(self):
# Create world
self.world = World()
# Create agents
self.tom = Agent('Tom')
self.world.addAgent(self.tom)
self.jerry = Agent('Jerry')
self.world.addAgent(self.jerry)
def addStates(self):
"""Create state features"""
self.world.defineState(self.tom.name,'health',int,lo=0,hi=100,
description='%s\'s wellbeing' % (self.tom.name))
self.world.setState(self.tom.name,'health',50)
self.world.defineState(self.jerry.name,'health',int,lo=0,hi=100,
description='%s\'s wellbeing' % (self.jerry.name))
self.world.setState(self.jerry.name,'health',50)
def addActions(self):
"""Create actions"""
self.chase = self.tom.addAction({'verb': 'chase','object': self.jerry.name})
self.hit = self.tom.addAction({'verb': 'hit','object': self.jerry.name})
self.run = self.jerry.addAction({'verb': 'run away'})
self.trick = self.jerry.addAction({'verb': 'trick','object': self.tom.name})
def addDynamics(self):
"""Create dynamics"""
tree = makeTree(incrementMatrix(stateKey(self.jerry.name,'health'),-10))
self.world.setDynamics(stateKey(self.jerry.name,'health'),self.hit,tree,enforceMin=True)
def addModels(self,rationality=1.):
self.tom.addModel('friend',rationality=rationality,parent=True)
self.tom.setReward(maximizeFeature(stateKey(self.jerry.name,'health')),1.,'friend')
self.tom.addModel('foe',rationality=rationality,parent=True)
self.tom.setReward(minimizeFeature(stateKey(self.jerry.name,'health')),1.,'foe')
def saveload(self):
"""Write scenario to file and then load from scratch"""
self.world.save('/tmp/psychsim_test.psy')
self.world = World('/tmp/psychsim_test.psy')
self.tom = self.world.agents[self.tom.name]
self.jerry = self.world.agents[self.jerry.name]
def testEnumeratedState(self):
self.addActions()
self.world.defineVariable(self.tom.name,ActionSet)
self.world.defineState(self.tom.name,'status',list,['dead','injured','healthy'])
self.world.setState(self.tom.name,'status','healthy')
goal = achieveFeatureValue(stateKey(self.tom.name,'status'),'healthy')
self.tom.setReward(goal,1.)
goal = achieveFeatureValue(stateKey(self.tom.name,'status'),'injured')
self.jerry.setReward(goal,1.)
self.saveload()
self.assertEqual(len(self.world.state[None]),1)
vector = self.world.state[None].domain()[0]
tVal = self.tom.reward(vector)
self.assertAlmostEqual(tVal,1.,8)
jVal = self.jerry.reward(vector)
self.assertAlmostEqual(jVal,0.,8)
for action in self.tom.actions:
encoding = self.world.value2float(self.tom.name,action)
self.assertEqual(action,self.world.float2value(self.tom.name,encoding))
def testBeliefModels(self):
self.addStates()
self.addActions()
self.addDynamics()
self.world.setOrder([self.tom.name])
self.tom.addModel('optimist')
self.tom.setBelief(stateKey(self.jerry.name,'health'),20,'optimist')
self.tom.addModel('pessimist')
self.world.setModel(self.jerry.name,True)
self.world.setMentalModel(self.jerry.name,self.tom.name,{'optimist': 0.5,'pessimist': 0.5})
actions = {self.tom.name: self.hit}
self.world.step(actions)
vector = self.world.state[None].domain()[0]
beliefs = self.jerry.getAttribute('beliefs',self.world.getModel(self.jerry.name,vector))
for belief in beliefs.domain():
model = self.world.getModel(self.tom.name,belief)
if self.tom.models[model].has_key('beliefs'):
nested = self.tom.models[model]['beliefs']
self.assertEqual(len(nested),1)
nested = nested.domain()[0]
self.assertEqual(len(nested),1)
self.assertAlmostEqual(nested[stateKey(self.jerry.name,'health')],10.,8)
def testObservation(self):
self.addStates()
self.addActions()
self.addDynamics()
self.world.setOrder([self.tom.name])
self.world.setModel(self.jerry.name,True)
key = stateKey(self.jerry.name,'health')
self.jerry.setBelief(key,Distribution({20: 0.5, 50: 0.5}))
tree = makeTree({'if': thresholdRow(key,40),
True: {'distribution': [(KeyedVector({CONSTANT: 50}),.8),
(KeyedVector({CONSTANT: 20}),.2)]},
False: {'distribution': [(KeyedVector({CONSTANT: 50}),.2),
(KeyedVector({CONSTANT: 20}),.8)]}})
self.jerry.defineObservation(key,tree)
actions = {self.tom.name: self.hit}
vector = self.world.state[None].domain()[0]
omegaDist = self.jerry.observe(vector,actions)
for omega in omegaDist.domain():
new = KeyedVector(vector)
model = self.jerry.index2model(self.jerry.stateEstimator(vector,new,omega))
beliefs = self.jerry.models[model]['beliefs']
if omega[key] > 30:
# We observed a high value, so we should have a stronger belief in the higher value
# which is now 40 after the hit
for belief in beliefs.domain():
if beliefs[belief] > 0.5:
self.assertAlmostEqual(belief[key],40,8)
else:
self.assertAlmostEqual(belief[key],10,8)
else:
# We observed a low value, so we should have a stronger belief in the lower value
# which is now 10 after the hit
for belief in beliefs.domain():
if beliefs[belief] < 0.5:
self.assertAlmostEqual(belief[key],40,8)
else:
self.assertAlmostEqual(belief[key],10,8)
def testUnobservedAction(self):
self.addStates()
self.addActions()
self.addDynamics()
self.addModels()
self.world.setOrder([self.tom.name])
self.world.setModel(self.jerry.name,True)
self.jerry.setBelief(stateKey(self.jerry.name,'health'),50)
self.world.setMentalModel(self.jerry.name,self.tom.name,{'friend': 0.5,'foe': 0.5})
tree = makeTree(True)
self.jerry.defineObservation(self.tom.name,tree,self.hit,domain=ActionSet)
tree = makeTree({'distribution': [(True,0.25),(False,0.75)]})
self.jerry.defineObservation(self.tom.name,tree,self.chase,domain=ActionSet)
vector = self.world.state[None].domain()[0]
self.saveload()
self.world.step({self.tom.name: self.hit})
vector = self.world.state[None].domain()[0]
def testRewardModels(self):
self.addStates()
self.addActions()
self.addDynamics()
self.addModels()
self.world.setOrder([self.tom.name])
# Add Jerry's model to the world (so that it gets updated)
self.world.setModel(self.jerry.name,True)
# Give Jerry uncertainty about Tom
self.world.setMentalModel(self.jerry.name,self.tom.name,{'friend': 0.5,'foe': 0.5})
self.saveload()
# Hitting should make Jerry think Tom is more of a foe
actions = {self.tom.name: self.hit}
self.world.step(actions)
vector = self.world.state[None].domain()[0]
belief01 = self.jerry.getAttribute('beliefs',self.world.getModel(self.jerry.name,vector))
key = modelKey(self.tom.name)
for belief in belief01.domain():
if self.tom.index2model(belief[key]) == 'foe':
prob01 = belief01[belief]
break
self.assertGreater(prob01,0.5)
# If we think of Tom as even more of an optimizer, then our update should be stronger
self.tom.setAttribute('rationality',10.,'foe')
self.tom.setAttribute('rationality',10.,'friend')
self.world.setMentalModel(self.jerry.name,self.tom.name,{'friend': 0.5,'foe': 0.5})
self.world.step(actions)
vector = self.world.state[None].domain()[0]
model = self.world.getModel(self.jerry.name,vector)
belief10 = self.jerry.getAttribute('beliefs',model)
key = modelKey(self.tom.name)
for belief in belief10.domain():
if self.tom.index2model(belief[key]) == 'foe':
prob10 = belief10[belief]
break
self.assertGreater(prob10,prob01)
# If we keep the same models, but get another observation, we should update even more
self.world.step(actions)
vector = self.world.state[None].domain()[0]
model = self.world.getModel(self.jerry.name,vector)
belief1010 = self.jerry.getAttribute('beliefs',model)
key = modelKey(self.tom.name)
for belief in belief1010.domain():
if self.tom.index2model(belief[key]) == 'foe':
prob1010 = belief1010[belief]
break
self.assertGreater(prob1010,prob10)
def testDynamics(self):
self.world.setOrder([self.tom.name])
self.addStates()
self.addActions()
self.addDynamics()
key = stateKey(self.jerry.name,'health')
self.assertEqual(len(self.world.state[None]),1)
vector = self.world.state[None].domain()[0]
self.assertTrue(vector.has_key(stateKey(self.tom.name,'health')))
self.assertTrue(vector.has_key(turnKey(self.tom.name)))
self.assertTrue(vector.has_key(key))
self.assertTrue(vector.has_key(CONSTANT))
self.assertEqual(len(vector),4)
self.assertEqual(vector[stateKey(self.tom.name,'health')],50)
self.assertEqual(vector[key],50)
outcome = self.world.step({self.tom.name: self.chase})
for i in range(7):
self.assertEqual(len(self.world.state[None]),1)
vector = self.world.state[None].domain()[0]
self.assertTrue(vector.has_key(stateKey(self.tom.name,'health')))
self.assertTrue(vector.has_key(turnKey(self.tom.name)))
self.assertTrue(vector.has_key(key))
self.assertTrue(vector.has_key(CONSTANT))
self.assertEqual(len(vector),4)
self.assertEqual(vector[stateKey(self.tom.name,'health')],50)
self.assertEqual(vector[key],max(50-10*i,0))
outcome = self.world.step({self.tom.name: self.hit})
self.saveload()
def testRewardOnOthers(self):
self.addStates()
self.addActions()
self.addDynamics()
self.world.setOrder([self.tom.name])
vector = self.world.state[None].domain()[0]
# Create Jerry's goals
goal = maximizeFeature(stateKey(self.jerry.name,'health'))
self.jerry.setReward(goal,1.)
jVal = -self.jerry.reward(vector)
# Create Tom's goals from scratch
minGoal = minimizeFeature(stateKey(self.jerry.name,'health'))
self.tom.setReward(minGoal,1.)
self.saveload()
tRawVal = self.tom.reward(vector)
self.assertAlmostEqual(jVal,tRawVal,8)
# Create Tom's goals as a function of Jerry's
self.tom.models[True]['R'].clear()
self.tom.setReward(self.jerry.name,-1.)
self.saveload()
tFuncVal = self.tom.reward(vector)
self.assertAlmostEqual(tRawVal,tFuncVal,8)
# Test effect of functional reward on value function
self.tom.setHorizon(1)
self.saveload()
vHit = self.tom.value(vector,self.hit)['V']
vChase = self.tom.value(vector,self.chase)['V']
self.assertAlmostEqual(vHit,vChase+.1,8)
def testReward(self):
self.addStates()
key = stateKey(self.jerry.name,'health')
goal = makeTree({'if': thresholdRow(key,5),
True: KeyedVector({key: -2}),
False: KeyedVector({key: -1})})
goal = goal.desymbolize(self.world.symbols)
self.jerry.setReward(goal,1.)
R = self.jerry.models[True]['R']
self.assertEqual(len(R),1)
newGoal = R.keys()[0]
self.assertEqual(newGoal,goal)
self.assertAlmostEqual(R[goal],1.,8)
self.jerry.setReward(goal,2.)
self.assertEqual(len(R),1)
self.assertEqual(R.keys()[0],goal)
self.assertAlmostEqual(R[goal],2.,8)
def testTurnDynamics(self):
self.addStates()
self.addActions()
self.world.setOrder([self.tom.name,self.jerry.name])
self.assertEqual(self.world.maxTurn,1)
self.saveload()
vector = self.world.state[None].domain()[0]
jTurn = turnKey(self.jerry.name)
tTurn = turnKey(self.tom.name)
self.assertEqual(self.world.next(),[self.tom.name])
self.assertEqual(vector[tTurn],0)
self.assertEqual(vector[jTurn],1)
self.world.step()
vector = self.world.state[None].domain()[0]
self.assertEqual(self.world.next(),[self.jerry.name])
self.assertEqual(vector[tTurn],1)
self.assertEqual(vector[jTurn],0)
self.world.step()
vector = self.world.state[None].domain()[0]
self.assertEqual(self.world.next(),[self.tom.name])
self.assertEqual(vector[tTurn],0)
self.assertEqual(vector[jTurn],1)
# Try some custom dynamics
self.world.setTurnDynamics(self.tom.name,self.hit,makeTree(noChangeMatrix(tTurn)))
self.world.setTurnDynamics(self.jerry.name,self.hit,makeTree(noChangeMatrix(tTurn)))
self.world.step()
vector = self.world.state[None].domain()[0]
self.assertEqual(self.world.next(),[self.tom.name])
self.assertEqual(vector[tTurn],0)
self.assertEqual(vector[jTurn],1)
self.world.step({self.tom.name: self.chase})
vector = self.world.state[None].domain()[0]
self.assertEqual(self.world.next(),[self.jerry.name])
self.assertEqual(vector[tTurn],1)
self.assertEqual(vector[jTurn],0)
def testStatic(self):
self.addStates()
self.addActions()
self.addDynamics()
self.addModels()
self.world.setModel(self.jerry.name,True)
self.world.setMentalModel(self.jerry.name,self.tom.name,{'friend': 0.5,'foe': 0.5})
self.world.setOrder([self.tom.name])
vector = self.world.state[None].domain()[0]
model = self.world.getModel(self.jerry.name,vector)
belief0 = self.jerry.models[model]['beliefs']
result = self.world.step({self.tom.name: self.hit})
vector = self.world.state[None].domain()[0]
model = self.world.getModel(self.jerry.name,vector)
belief1 = self.jerry.models[model]['beliefs']
key = modelKey(self.tom.name)
for vector in belief0.domain():
if self.tom.index2model(vector[key]) == 'friend':
self.assertGreater(belief0[vector],belief1[vector])
else:
self.assertGreater(belief1[vector],belief0[vector])
# Now with the static beliefs
self.jerry.setAttribute('static',True,model)
self.saveload()
self.world.step()
vector = self.world.state[None].domain()[0]
model = self.world.getModel(self.jerry.name,vector)
belief2 = self.jerry.models[model]['beliefs']
for vector in belief1.domain():
self.assertAlmostEqual(belief1[vector],belief2[vector],8)
def scenarioCreationUseCase(enemy='Sylvania',model='powell',web=False,
fCollapse=None,sCollapse=None,maxRounds=15):
"""
An example of how to create a scenario
@param enemy: the name of the agent-controlled side, i.e., Freedonia's opponent (default: Sylvania)
@type enemy: str
@param model: which model do we use (default is "powell")
@type model: powell or slantchev
@param web: if C{True}, then create the web-based experiment scenario (default: C{False})
@type web: bool
@param fCollapse: the probability that Freedonia collapses (under powell, default: 0.1) or loses battle (under slantchev, default: 0.7)
@type fCollapse: float
@param sCollapse: the probability that Sylvania collapses, under powell (default: 0.1)
@type sCollapse: float
@param maxRounds: the maximum number of game rounds (default: 15)
@type maxRounds: int
@return: the scenario created
@rtype: L{World}
"""
# Handle defaults for battle probabilities, under each model
posLo = 0
posHi = 10
if fCollapse is None:
if model == 'powell':
fCollapse = 0.1
elif model == 'slantchev':
fCollapse = 0.7
if sCollapse is None:
sCollapse = 0.1
# Create scenario
world = World()
# Agents
free = Agent('Freedonia')
world.addAgent(free)
sylv = Agent(enemy)
world.addAgent(sylv)
# User state
world.defineState(free.name,'troops',int,lo=0,hi=50000,
description='Number of troops you have left')
free.setState('troops',40000)
world.defineState(free.name,'territory',int,lo=0,hi=100,
description='Percentage of disputed territory owned by you')
free.setState('territory',15)
world.defineState(free.name,'cost',int,lo=0,hi=50000,
description='Number of troops %s loses in an attack' % (free.name))
free.setState('cost',2000)
world.defineState(free.name,'position',int,lo=posLo,hi=posHi,
description='Current status of war (%d=%s is winner, %d=you are winner)' % (posLo,sylv.name,posHi))
free.setState('position',5)
world.defineState(free.name,'offered',int,lo=0,hi=100,
description='Percentage of disputed territory that %s last offered to you' % (sylv.name))
free.setState('offered',0)
if model == 'slantchev':
# Compute new value for territory only *after* computing new value for position
world.addDependency(stateKey(free.name,'territory'),stateKey(free.name,'position'))
# Agent state
world.defineState(sylv.name,'troops',int,lo=0,hi=500000,
description='Number of troops %s has left' % (sylv.name))
sylv.setState('troops',30000)
world.defineState(sylv.name,'cost',int,lo=0,hi=50000,
description='Number of troops %s loses in an attack' % (sylv.name))
sylv.setState('cost',2000)
world.defineState(sylv.name,'offered',int,lo=0,hi=100,
description='Percentage of disputed territory that %s last offered to %s' % (free.name,sylv.name))
sylv.setState('offered',0)
# World state
world.defineState(None,'treaty',bool,
description='Have the two sides reached an agreement?')
world.setState(None,'treaty',False)
# Stage of negotiation, illustrating the use of an enumerated state feature
world.defineState(None,'phase',list,['offer','respond','rejection','end','paused','engagement'],
description='The current stage of the negotiation game')
world.setState(None,'phase','paused')
# Game model, static descriptor
world.defineState(None,'model',list,['powell','slantchev'],
description='The model underlying the negotiation game')
world.setState(None,'model',model)
# Round of negotiation
world.defineState(None,'round',int,description='The current round of the negotiation')
world.setState(None,'round',0)
if not web:
# Relationship value
key = world.defineRelation(free.name,sylv.name,'trusts')
world.setFeature(key,0.)
# Game over if there is a treaty
world.addTermination(makeTree({'if': trueRow(stateKey(None,'treaty')),
True: True, False: False}))
# Game over if Freedonia has no territory
world.addTermination(makeTree({'if': thresholdRow(stateKey(free.name,'territory'),1),
True: False, False: True}) )
# Game over if Freedonia has all the territory
world.addTermination(makeTree({'if': thresholdRow(stateKey(free.name,'territory'),99),
True: True, False: False}))
# Game over if number of rounds exceeds limit
world.addTermination(makeTree({'if': thresholdRow(stateKey(None,'round'),maxRounds),
True: True, False: False}))
# Turn order: Uncomment the following if you want agents to act in parallel
# world.setOrder([set(world.agents.keys())])
# Turn order: Uncomment the following if you want agents to act sequentially
world.setOrder([free.name,sylv.name])
# User actions
freeBattle = free.addAction({'verb': 'attack','object': sylv.name})
for amount in range(20,100,20):
free.addAction({'verb': 'offer','object': sylv.name,'amount': amount})
if model == 'powell':
# Powell has null stages
freeNOP = free.addAction({'verb': 'continue'})
elif model == 'slantchev':
# Slantchev has both sides receiving offers
free.addAction({'verb': 'accept offer','object': sylv.name})
free.addAction({'verb': 'reject offer','object': sylv.name})
# Agent actions
sylvBattle = sylv.addAction({'verb': 'attack','object': free.name})
sylvAccept = sylv.addAction({'verb': 'accept offer','object': free.name})
sylvReject = sylv.addAction({'verb': 'reject offer','object': free.name})
if model == 'powell':
# Powell has null stages
sylvNOP = sylv.addAction({'verb': 'continue'})
elif model == 'slantchev':
# Slantchev has both sides making offers
for amount in range(10,100,10):
sylv.addAction({'verb': 'offer','object': free.name,'amount': amount})
# Restrictions on when actions are legal, based on phase of game
for action in filterActions({'verb': 'offer'},free.actions | sylv.actions):
agent = world.agents[action['subject']]
agent.setLegal(action,makeTree({'if': equalRow(stateKey(None,'phase'),'offer'),
True: True, # Offers are legal in the offer phase
False: False})) # Offers are illegal in all other phases
if model == 'powell':
# Powell has a special rejection phase
for action in [freeNOP,freeBattle]:
free.setLegal(action,makeTree({'if': equalRow(stateKey(None,'phase'),'rejection'),
True: True, # Attacking and doing nothing are legal only in rejection phase
False: False})) # Attacking and doing nothing are illegal in all other phases
# Once offered, agent can respond
if model == 'powell':
# Under Powell, only Sylvania has to respond, and it can attack
responses = [sylvBattle,sylvAccept,sylvReject]
elif model == 'slantchev':
# Under Slantchev, only accept/reject
responses = filterActions({'verb': 'accept offer'},free.actions | sylv.actions)
responses += filterActions({'verb': 'reject offer'},free.actions | sylv.actions)
for action in responses:
agent = world.agents[action['subject']]
agent.setLegal(action,makeTree({'if': equalRow(stateKey(None,'phase'),'respond'),
True: True, # Offeree must act in the response phase
False: False})) # Offeree cannot act in any other phase
if model == 'powell':
# NOP is legal in exactly opposite situations to all other actions
sylv.setLegal(sylvNOP,makeTree({'if': equalRow(stateKey(None,'phase'),'end'),
True: True, # Sylvania does not do anything in the null phase after Freedonia responds to rejection
False: False})) # Sylvania must act in its other phases
if model == 'slantchev':
# Attacking legal only under engagement phase
for action in filterActions({'verb': 'attack'},free.actions | sylv.actions):
agent = world.agents[action['subject']]
agent.setLegal(action,makeTree({'if': equalRow(stateKey(None,'phase'),'engagement'),
True: True, # Attacking legal only in engagement
False: False})) # Attacking legal every other phase
# Goals for Freedonia
goalFTroops = maximizeFeature(stateKey(free.name,'troops'))
free.setReward(goalFTroops,1.)
goalFTerritory = maximizeFeature(stateKey(free.name,'territory'))
free.setReward(goalFTerritory,1.)
# Goals for Sylvania
goalSTroops = maximizeFeature(stateKey(sylv.name,'troops'))
sylv.setReward(goalSTroops,1.)
goalSTerritory = minimizeFeature(stateKey(free.name,'territory'))
sylv.setReward(goalSTerritory,1.)
# Possible goals applicable to both
goalAgreement = maximizeFeature(stateKey(None,'treaty'))
# Silly goal, provided as an example of an achievement goal
goalAchieve = achieveFeatureValue(stateKey(None,'phase'),'respond')
# Horizons
if model == 'powell':
free.setAttribute('horizon',4)
sylv.setAttribute('horizon',4)
elif model == 'slantchev':
free.setAttribute('horizon',6)
sylv.setAttribute('horizon',6)
# Discount factors
free.setAttribute('discount',-1)
sylv.setAttribute('discount',-1)
# Levels of belief
free.setRecursiveLevel(2)
sylv.setRecursiveLevel(2)
# Dynamics of battle
freeTroops = stateKey(free.name,'troops')
freeTerr = stateKey(free.name,'territory')
sylvTroops = stateKey(sylv.name,'troops')
# Effect of fighting
for action in filterActions({'verb': 'attack'},free.actions | sylv.actions):
# Effect on troops (cost of battle)
tree = makeTree(addFeatureMatrix(freeTroops,stateKey(free.name,'cost'),-1.))
world.setDynamics(freeTroops,action,tree,enforceMin=not web)
tree = makeTree(addFeatureMatrix(sylvTroops,stateKey(sylv.name,'cost'),-1.))
world.setDynamics(sylvTroops,action,tree,enforceMin=not web)
if model == 'powell':
# Effect on territory (probability of collapse)
tree = makeTree({'distribution': [
({'distribution': [(setToConstantMatrix(freeTerr,100),1.-fCollapse), # Sylvania collapses, Freedonia does not
(noChangeMatrix(freeTerr), fCollapse)]}, # Both collapse
sCollapse),
({'distribution': [(setToConstantMatrix(freeTerr,0),fCollapse), # Freedonia collapses, Sylvania does not
(noChangeMatrix(freeTerr), 1.-fCollapse)]}, # Neither collapses
1.-sCollapse)]})
world.setDynamics(freeTerr,action,tree)
elif model == 'slantchev':
# Effect on position
pos = stateKey(free.name,'position')
tree = makeTree({'distribution': [(incrementMatrix(pos,1),1.-fCollapse), # Freedonia wins battle
(incrementMatrix(pos,-1),fCollapse)]}) # Freedonia loses battle
world.setDynamics(pos,action,tree)
# Effect on territory
tree = makeTree({'if': thresholdRow(pos,posHi-.5),
True: setToConstantMatrix(freeTerr,100), # Freedonia won
False: {'if': thresholdRow(pos,posLo+.5),
True: noChangeMatrix(freeTerr),
False: setToConstantMatrix(freeTerr,0)}}) # Freedonia lost
world.setDynamics(freeTerr,action,tree)
# Dynamics of offers
for index in range(2):
atom = Action({'subject': world.agents.keys()[index],'verb': 'offer',
'object': world.agents.keys()[1-index]})
if atom['subject'] == free.name or model != 'powell':
offer = stateKey(atom['object'],'offered')
amount = actionKey('amount')
tree = makeTree({'if': trueRow(stateKey(None,'treaty')),
True: noChangeMatrix(offer),
False: setToConstantMatrix(offer,amount)})
world.setDynamics(offer,atom,tree,enforceMax=not web)
# Dynamics of treaties
for action in filterActions({'verb': 'accept offer'},free.actions | sylv.actions):
# Accepting an offer means that there is now a treaty
key = stateKey(None,'treaty')
tree = makeTree(setTrueMatrix(key))
world.setDynamics(key,action,tree)
# Accepting offer sets territory
offer = stateKey(action['subject'],'offered')
territory = stateKey(free.name,'territory')
if action['subject'] == free.name:
# Freedonia accepts sets territory to last offer
tree = makeTree(setToFeatureMatrix(territory,offer))
world.setDynamics(freeTerr,action,tree)
else:
# Sylvania accepts sets territory to 1-last offer
tree = makeTree(setToFeatureMatrix(territory,offer,pct=-1.,shift=100.))
world.setDynamics(freeTerr,action,tree)
# Dynamics of phase
phase = stateKey(None,'phase')
roundKey = stateKey(None,'round')
# OFFER -> RESPOND
for index in range(2):
action = Action({'subject': world.agents.keys()[index],'verb': 'offer',
'object': world.agents.keys()[1-index]})
if action['subject'] == free.name or model != 'powell':
tree = makeTree(setToConstantMatrix(phase,'respond'))
world.setDynamics(phase,action,tree)
# RESPOND -> REJECTION or ENGAGEMENT
for action in filterActions({'verb': 'reject offer'},free.actions | sylv.actions):
if model == 'powell':
tree = makeTree(setToConstantMatrix(phase,'rejection'))
elif model == 'slantchev':
tree = makeTree(setToConstantMatrix(phase,'engagement'))
world.setDynamics(phase,action,tree)
# accepting -> OFFER
for action in filterActions({'verb': 'accept offer'},free.actions | sylv.actions):
tree = makeTree(setToConstantMatrix(phase,'offer'))
world.setDynamics(phase,action,tree)
# attacking -> OFFER
for action in filterActions({'verb': 'attack'},free.actions | sylv.actions):
tree = makeTree(setToConstantMatrix(phase,'offer'))
world.setDynamics(phase,action,tree)
if action['subject'] == sylv.name or model == 'slantchev':
tree = makeTree(incrementMatrix(roundKey,1))
world.setDynamics(roundKey,action,tree)
if model == 'powell':
# REJECTION -> END
for atom in [freeNOP,freeBattle]:
tree = makeTree(setToConstantMatrix(phase,'end'))
world.setDynamics(phase,atom,tree)
# END -> OFFER
atom = Action({'subject': sylv.name,'verb': 'continue'})
tree = makeTree(setToConstantMatrix(phase,'offer'))
world.setDynamics(phase,atom,tree)
tree = makeTree(incrementMatrix(roundKey,1))
world.setDynamics(roundKey,atom,tree)
if not web:
# Relationship dynamics: attacking is bad for trust
atom = Action({'subject': sylv.name,'verb': 'attack','object': free.name})
key = binaryKey(free.name,sylv.name,'trusts')
tree = makeTree(approachMatrix(key,0.1,-1.))
world.setDynamics(key,atom,tree)
# Handcrafted policy for Freedonia
# free.setPolicy(makeTree({'if': equalRow('phase','respond'),
# # Accept an offer greater than 50
# True: {'if': thresholdRow(stateKey(free.name,'offered'),50),
# True: Action({'subject': free.name,'verb': 'accept offer','object': sylv.name}),
# False: Action({'subject': free.name,'verb': 'reject offer','object': sylv.name})},
# False: {'if': equalRow('phase','engagement'),
# # Attack during engagement phase
# True: Action({'subject': free.name,'verb': 'attack','object': sylv.name}),
# # Agent decides how what to do otherwise
# False: False}}))
# Mental models of enemy
# Example of creating a model with incorrect reward all at once (a version of Freedonia who cares about reaching agreement as well)
# sylv.addModel('false',R={goalSTroops: 10.,goalSTerritory: 1.,goalAgreement: 1.},
# rationality=1.,selection='distribution',parent=True)
# Example of creating a model with incorrect beliefs
sylv.addModel('false',rationality=10.,selection='distribution',parent=True)
key = stateKey(free.name,'position')
# Sylvania believes position to be fixed at 3
sylv.setBelief(key,3,'false')
# Freedonia is truly unsure about position (50% chance of being 7, 50% of being 3)
world.setModel(free.name,True)
free.setBelief(key,Distribution({7: 0.5,3: 0.5}),True)
# Observations about military position
tree = makeTree({'if': thresholdRow(key,1),
True: {'if': thresholdRow(key,9),
True: {'distribution': [(KeyedVector({key: 1}),0.9),
(KeyedVector({key: 1,CONSTANT: -1}),0.1)]},
False: {'distribution': [(KeyedVector({key: 1}),0.8),
(KeyedVector({key: 1,CONSTANT: -1}),0.1),
(KeyedVector({key: 1,CONSTANT: 1}),0.1)]}},
False: {'distribution': [(KeyedVector({key: 1}),0.9),
(KeyedVector({key: 1,CONSTANT: 1}),0.1)]}})
free.defineObservation(key,tree)
# Example of setting model parameters separately
sylv.addModel('true',parent=True)
sylv.setAttribute('rationality',10.,'true') # Override real agent's rationality with this value
sylv.setAttribute('selection','distribution','true')
world.setMentalModel(free.name,sylv.name,{'false': 0.9,'true': 0.1})
# Goal of fooling Sylvania
goalDeception = achieveFeatureValue(modelKey(sylv.name),sylv.model2index('false'))
return world
def scenarioCreationUseCase(enemy='Sylvania',
model='powell',
web=False,
fCollapse=None,
sCollapse=None,
maxRounds=15):
"""
An example of how to create a scenario
@param enemy: the name of the agent-controlled side, i.e., Freedonia's opponent (default: Sylvania)
@type enemy: str
@param model: which model do we use (default is "powell")
@type model: powell or slantchev
@param web: if C{True}, then create the web-based experiment scenario (default: C{False})
@type web: bool
@param fCollapse: the probability that Freedonia collapses (under powell, default: 0.1) or loses battle (under slantchev, default: 0.7)
@type fCollapse: float
@param sCollapse: the probability that Sylvania collapses, under powell (default: 0.1)
@type sCollapse: float
@param maxRounds: the maximum number of game rounds (default: 15)
@type maxRounds: int
@return: the scenario created
@rtype: L{World}
"""
# Handle defaults for battle probabilities, under each model
posLo = 0
posHi = 10
if fCollapse is None:
if model == 'powell':
fCollapse = 0.1
elif model == 'slantchev':
fCollapse = 0.7
if sCollapse is None:
sCollapse = 0.1
# Create scenario
world = World()
# Agents
free = Agent('Freedonia')
world.addAgent(free)
sylv = Agent(enemy)
world.addAgent(sylv)
# User state
world.defineState(free.name,
'troops',
int,
lo=0,
hi=50000,
description='Number of troops you have left')
free.setState('troops', 40000)
world.defineState(
free.name,
'territory',
int,
lo=0,
hi=100,
description='Percentage of disputed territory owned by you')
free.setState('territory', 15)
world.defineState(free.name,
'cost',
int,
lo=0,
hi=50000,
description='Number of troops %s loses in an attack' %
(free.name))
free.setState('cost', 2000)
world.defineState(
free.name,
'position',
int,
lo=posLo,
hi=posHi,
description='Current status of war (%d=%s is winner, %d=you are winner)'
% (posLo, sylv.name, posHi))
free.setState('position', 5)
world.defineState(
free.name,
'offered',
int,
lo=0,
hi=100,
description=
'Percentage of disputed territory that %s last offered to you' %
(sylv.name))
free.setState('offered', 0)
if model == 'slantchev':
# Compute new value for territory only *after* computing new value for position
world.addDependency(stateKey(free.name, 'territory'),
stateKey(free.name, 'position'))
# Agent state
world.defineState(sylv.name,
'troops',
int,
lo=0,
hi=500000,
description='Number of troops %s has left' % (sylv.name))
sylv.setState('troops', 30000)
world.defineState(sylv.name,
'cost',
int,
lo=0,
hi=50000,
description='Number of troops %s loses in an attack' %
(sylv.name))
sylv.setState('cost', 2000)
world.defineState(
sylv.name,
'offered',
int,
lo=0,
hi=100,
description=
'Percentage of disputed territory that %s last offered to %s' %
(free.name, sylv.name))
sylv.setState('offered', 0)
# World state
world.defineState(None,
'treaty',
bool,
description='Have the two sides reached an agreement?')
world.setState(None, 'treaty', False)
# Stage of negotiation, illustrating the use of an enumerated state feature
world.defineState(
None,
'phase',
list, ['offer', 'respond', 'rejection', 'end', 'paused', 'engagement'],
description='The current stage of the negotiation game')
world.setState(None, 'phase', 'paused')
# Game model, static descriptor
world.defineState(None,
'model',
list, ['powell', 'slantchev'],
description='The model underlying the negotiation game')
world.setState(None, 'model', model)
# Round of negotiation
world.defineState(None,
'round',
int,
description='The current round of the negotiation')
world.setState(None, 'round', 0)
if not web:
# Relationship value
key = world.defineRelation(free.name, sylv.name, 'trusts')
world.setFeature(key, 0.)
# Game over if there is a treaty
world.addTermination(
makeTree({
'if': trueRow(stateKey(None, 'treaty')),
True: True,
False: False
}))
# Game over if Freedonia has no territory
world.addTermination(
makeTree({
'if': thresholdRow(stateKey(free.name, 'territory'), 1),
True: False,
False: True
}))
# Game over if Freedonia has all the territory
world.addTermination(
makeTree({
'if': thresholdRow(stateKey(free.name, 'territory'), 99),
True: True,
False: False
}))
# Game over if number of rounds exceeds limit
world.addTermination(
makeTree({
'if': thresholdRow(stateKey(None, 'round'), maxRounds),
True: True,
False: False
}))
# Turn order: Uncomment the following if you want agents to act in parallel
# world.setOrder([set(world.agents.keys())])
# Turn order: Uncomment the following if you want agents to act sequentially
world.setOrder([free.name, sylv.name])
# User actions
freeBattle = free.addAction({'verb': 'attack', 'object': sylv.name})
for amount in range(20, 100, 20):
free.addAction({
'verb': 'offer',
'object': sylv.name,
'amount': amount
})
if model == 'powell':
# Powell has null stages
freeNOP = free.addAction({'verb': 'continue'})
elif model == 'slantchev':
# Slantchev has both sides receiving offers
free.addAction({'verb': 'accept offer', 'object': sylv.name})
free.addAction({'verb': 'reject offer', 'object': sylv.name})
# Agent actions
sylvBattle = sylv.addAction({'verb': 'attack', 'object': free.name})
sylvAccept = sylv.addAction({'verb': 'accept offer', 'object': free.name})
sylvReject = sylv.addAction({'verb': 'reject offer', 'object': free.name})
if model == 'powell':
# Powell has null stages
sylvNOP = sylv.addAction({'verb': 'continue'})
elif model == 'slantchev':
# Slantchev has both sides making offers
for amount in range(10, 100, 10):
sylv.addAction({
'verb': 'offer',
'object': free.name,
'amount': amount
})
# Restrictions on when actions are legal, based on phase of game
for action in filterActions({'verb': 'offer'},
free.actions | sylv.actions):
agent = world.agents[action['subject']]
agent.setLegal(
action,
makeTree({
'if': equalRow(stateKey(None, 'phase'), 'offer'),
True: True, # Offers are legal in the offer phase
False: False
})) # Offers are illegal in all other phases
if model == 'powell':
# Powell has a special rejection phase
for action in [freeNOP, freeBattle]:
free.setLegal(
action,
makeTree({
'if': equalRow(stateKey(None, 'phase'), 'rejection'),
True:
True, # Attacking and doing nothing are legal only in rejection phase
False: False
})
) # Attacking and doing nothing are illegal in all other phases
# Once offered, agent can respond
if model == 'powell':
# Under Powell, only Sylvania has to respond, and it can attack
responses = [sylvBattle, sylvAccept, sylvReject]
elif model == 'slantchev':
# Under Slantchev, only accept/reject
responses = filterActions({'verb': 'accept offer'},
free.actions | sylv.actions)
responses += filterActions({'verb': 'reject offer'},
free.actions | sylv.actions)
for action in responses:
agent = world.agents[action['subject']]
agent.setLegal(
action,
makeTree({
'if': equalRow(stateKey(None, 'phase'), 'respond'),
True: True, # Offeree must act in the response phase
False: False
})) # Offeree cannot act in any other phase
if model == 'powell':
# NOP is legal in exactly opposite situations to all other actions
sylv.setLegal(
sylvNOP,
makeTree({
'if': equalRow(stateKey(None, 'phase'), 'end'),
True:
True, # Sylvania does not do anything in the null phase after Freedonia responds to rejection
False: False
})) # Sylvania must act in its other phases
if model == 'slantchev':
# Attacking legal only under engagement phase
for action in filterActions({'verb': 'attack'},
free.actions | sylv.actions):
agent = world.agents[action['subject']]
agent.setLegal(
action,
makeTree({
'if': equalRow(stateKey(None, 'phase'), 'engagement'),
True: True, # Attacking legal only in engagement
False: False
})) # Attacking legal every other phase
# Goals for Freedonia
goalFTroops = maximizeFeature(stateKey(free.name, 'troops'))
free.setReward(goalFTroops, 1.)
goalFTerritory = maximizeFeature(stateKey(free.name, 'territory'))
free.setReward(goalFTerritory, 1.)
# Goals for Sylvania
goalSTroops = maximizeFeature(stateKey(sylv.name, 'troops'))
sylv.setReward(goalSTroops, 1.)
goalSTerritory = minimizeFeature(stateKey(free.name, 'territory'))
sylv.setReward(goalSTerritory, 1.)
# Possible goals applicable to both
goalAgreement = maximizeFeature(stateKey(None, 'treaty'))
# Silly goal, provided as an example of an achievement goal
goalAchieve = achieveFeatureValue(stateKey(None, 'phase'), 'respond')
# Horizons
if model == 'powell':
free.setAttribute('horizon', 4)
sylv.setAttribute('horizon', 4)
elif model == 'slantchev':
free.setAttribute('horizon', 6)
sylv.setAttribute('horizon', 6)
# Discount factors
free.setAttribute('discount', -1)
sylv.setAttribute('discount', -1)
# Levels of belief
free.setRecursiveLevel(2)
sylv.setRecursiveLevel(2)
# Dynamics of battle
freeTroops = stateKey(free.name, 'troops')
freeTerr = stateKey(free.name, 'territory')
sylvTroops = stateKey(sylv.name, 'troops')
# Effect of fighting
for action in filterActions({'verb': 'attack'},
free.actions | sylv.actions):
# Effect on troops (cost of battle)
tree = makeTree(
addFeatureMatrix(freeTroops, stateKey(free.name, 'cost'), -1.))
world.setDynamics(freeTroops, action, tree, enforceMin=not web)
tree = makeTree(
addFeatureMatrix(sylvTroops, stateKey(sylv.name, 'cost'), -1.))
world.setDynamics(sylvTroops, action, tree, enforceMin=not web)
if model == 'powell':
# Effect on territory (probability of collapse)
tree = makeTree({
'distribution': [
(
{
'distribution': [
(setToConstantMatrix(freeTerr,
100), 1. - fCollapse
), # Sylvania collapses, Freedonia does not
(noChangeMatrix(freeTerr), fCollapse)
]
}, # Both collapse
sCollapse),
(
{
'distribution': [
(setToConstantMatrix(freeTerr, 0), fCollapse
), # Freedonia collapses, Sylvania does not
(noChangeMatrix(freeTerr), 1. - fCollapse)
]
}, # Neither collapses
1. - sCollapse)
]
})
world.setDynamics(freeTerr, action, tree)
elif model == 'slantchev':
# Effect on position
pos = stateKey(free.name, 'position')
tree = makeTree({
'distribution': [
(incrementMatrix(pos, 1),
1. - fCollapse), # Freedonia wins battle
(incrementMatrix(pos, -1), fCollapse)
]
}) # Freedonia loses battle
world.setDynamics(pos, action, tree)
# Effect on territory
tree = makeTree({
'if': thresholdRow(pos, posHi - .5),
True: setToConstantMatrix(freeTerr, 100), # Freedonia won
False: {
'if': thresholdRow(pos, posLo + .5),
True: noChangeMatrix(freeTerr),
False: setToConstantMatrix(freeTerr, 0)
}
}) # Freedonia lost
world.setDynamics(freeTerr, action, tree)
# Dynamics of offers
for index in range(2):
atom = Action({
'subject': world.agents.keys()[index],
'verb': 'offer',
'object': world.agents.keys()[1 - index]
})
if atom['subject'] == free.name or model != 'powell':
offer = stateKey(atom['object'], 'offered')
amount = actionKey('amount')
tree = makeTree({
'if': trueRow(stateKey(None, 'treaty')),
True: noChangeMatrix(offer),
False: setToConstantMatrix(offer, amount)
})
world.setDynamics(offer, atom, tree, enforceMax=not web)
# Dynamics of treaties
for action in filterActions({'verb': 'accept offer'},
free.actions | sylv.actions):
# Accepting an offer means that there is now a treaty
key = stateKey(None, 'treaty')
tree = makeTree(setTrueMatrix(key))
world.setDynamics(key, action, tree)
# Accepting offer sets territory
offer = stateKey(action['subject'], 'offered')
territory = stateKey(free.name, 'territory')
if action['subject'] == free.name:
# Freedonia accepts sets territory to last offer
tree = makeTree(setToFeatureMatrix(territory, offer))
world.setDynamics(freeTerr, action, tree)
else:
# Sylvania accepts sets territory to 1-last offer
tree = makeTree(
setToFeatureMatrix(territory, offer, pct=-1., shift=100.))
world.setDynamics(freeTerr, action, tree)
# Dynamics of phase
phase = stateKey(None, 'phase')
roundKey = stateKey(None, 'round')
# OFFER -> RESPOND
for index in range(2):
action = Action({
'subject': world.agents.keys()[index],
'verb': 'offer',
'object': world.agents.keys()[1 - index]
})
if action['subject'] == free.name or model != 'powell':
tree = makeTree(setToConstantMatrix(phase, 'respond'))
world.setDynamics(phase, action, tree)
# RESPOND -> REJECTION or ENGAGEMENT
for action in filterActions({'verb': 'reject offer'},
free.actions | sylv.actions):
if model == 'powell':
tree = makeTree(setToConstantMatrix(phase, 'rejection'))
elif model == 'slantchev':
tree = makeTree(setToConstantMatrix(phase, 'engagement'))
world.setDynamics(phase, action, tree)
# accepting -> OFFER
for action in filterActions({'verb': 'accept offer'},
free.actions | sylv.actions):
tree = makeTree(setToConstantMatrix(phase, 'offer'))
world.setDynamics(phase, action, tree)
# attacking -> OFFER
for action in filterActions({'verb': 'attack'},
free.actions | sylv.actions):
tree = makeTree(setToConstantMatrix(phase, 'offer'))
world.setDynamics(phase, action, tree)
if action['subject'] == sylv.name or model == 'slantchev':
tree = makeTree(incrementMatrix(roundKey, 1))
world.setDynamics(roundKey, action, tree)
if model == 'powell':
# REJECTION -> END
for atom in [freeNOP, freeBattle]:
tree = makeTree(setToConstantMatrix(phase, 'end'))
world.setDynamics(phase, atom, tree)
# END -> OFFER
atom = Action({'subject': sylv.name, 'verb': 'continue'})
tree = makeTree(setToConstantMatrix(phase, 'offer'))
world.setDynamics(phase, atom, tree)
tree = makeTree(incrementMatrix(roundKey, 1))
world.setDynamics(roundKey, atom, tree)
if not web:
# Relationship dynamics: attacking is bad for trust
atom = Action({
'subject': sylv.name,
'verb': 'attack',
'object': free.name
})
key = binaryKey(free.name, sylv.name, 'trusts')
tree = makeTree(approachMatrix(key, 0.1, -1.))
world.setDynamics(key, atom, tree)
# Handcrafted policy for Freedonia
# free.setPolicy(makeTree({'if': equalRow('phase','respond'),
# # Accept an offer greater than 50
# True: {'if': thresholdRow(stateKey(free.name,'offered'),50),
# True: Action({'subject': free.name,'verb': 'accept offer','object': sylv.name}),
# False: Action({'subject': free.name,'verb': 'reject offer','object': sylv.name})},
# False: {'if': equalRow('phase','engagement'),
# # Attack during engagement phase
# True: Action({'subject': free.name,'verb': 'attack','object': sylv.name}),
# # Agent decides how what to do otherwise
# False: False}}))
# Mental models of enemy
# Example of creating a model with incorrect reward all at once (a version of Freedonia who cares about reaching agreement as well)
# sylv.addModel('false',R={goalSTroops: 10.,goalSTerritory: 1.,goalAgreement: 1.},
# rationality=1.,selection='distribution',parent=True)
# Example of creating a model with incorrect beliefs
sylv.addModel('false',
rationality=10.,
selection='distribution',
parent=True)
key = stateKey(free.name, 'position')
# Sylvania believes position to be fixed at 3
sylv.setBelief(key, 3, 'false')
# Freedonia is truly unsure about position (50% chance of being 7, 50% of being 3)
world.setModel(free.name, True)
free.setBelief(key, Distribution({7: 0.5, 3: 0.5}), True)
# Observations about military position
tree = makeTree({
'if': thresholdRow(key, 1),
True: {
'if': thresholdRow(key, 9),
True: {
'distribution': [(KeyedVector({key: 1}), 0.9),
(KeyedVector({
key: 1,
CONSTANT: -1
}), 0.1)]
},
False: {
'distribution': [(KeyedVector({key: 1}), 0.8),
(KeyedVector({
key: 1,
CONSTANT: -1
}), 0.1),
(KeyedVector({
key: 1,
CONSTANT: 1
}), 0.1)]
}
},
False: {
'distribution': [(KeyedVector({key: 1}), 0.9),
(KeyedVector({
key: 1,
CONSTANT: 1
}), 0.1)]
}
})
free.defineObservation(key, tree)
# Example of setting model parameters separately
sylv.addModel('true', parent=True)
sylv.setAttribute(
'rationality', 10.,
'true') # Override real agent's rationality with this value
sylv.setAttribute('selection', 'distribution', 'true')
world.setMentalModel(free.name, sylv.name, {'false': 0.9, 'true': 0.1})
# Goal of fooling Sylvania
goalDeception = achieveFeatureValue(modelKey(sylv.name),
sylv.model2index('false'))
return world
if name == 'survival':
tree = {'if': equalRow(stateKey(resident.name,'location',True),'Seattle'),
True: tree, False: setTrueMatrix(objective['key'])}
world.setDynamics(objective['key'],option,makeTree(tree))
# Movement dynamics
world.setDynamics(location,behaviors['leave']['action'],makeTree(setToConstantMatrix(location,'beyond')))
world.setDynamics(location,behaviors['return']['action'],makeTree(setToConstantMatrix(location,'Seattle')))
# Phase dynamics
world.setDynamics('phase',True,makeTree({'if': equalRow('phase','where'),
True: setToConstantMatrix('phase','how'),
False: setToConstantMatrix('phase','where')}))
# Decision-making parameters
resident.setAttribute('horizon',2)
resident.setAttribute('selection','distribution')
world.save('anthrax.psy')
# world.printState()
for tree,weight in resident.getAttribute('R').items():
print weight,tree
decision = Distribution()
for vector in world.state[None].domain():
world.printVector(vector)
result = resident.decide(vector,selection='distribution')
for action in result['action'].domain():
decision.addProb(action,world.state[None][vector]*result['action'][action])
print 'Choice:',action
if __name__ == '__main__':
# sets up log to screen
logging.basicConfig(format='%(message)s',
level=logging.DEBUG if DEBUG else logging.INFO)
# create world and add agents
world = World()
ag_producer = Agent('Producer')
world.addAgent(ag_producer)
ag_consumer = Agent('Consumer')
world.addAgent(ag_consumer)
agents = [ag_producer, ag_consumer]
# agent settings
ag_producer.setAttribute('discount', 1)
ag_producer.setHorizon(HORIZON)
ag_consumer.setAttribute('discount', 1)
ag_consumer.setHorizon(HORIZON)
# add variables (capacity and asked/received amounts)
var_half_cap = world.defineState(ag_producer.name, 'half capacity', bool)
world.setFeature(var_half_cap, False)
var_ask_amnt = world.defineState(ag_producer.name,
'asked amount',
int,
lo=0,
hi=100)
world.setFeature(var_ask_amnt, 0)
var_rcv_amnt = world.defineState(ag_consumer.name,
'received amount',
'i.e., in its mind the position is changing, but is not aligned with the real/true position.'
# parameters
HORIZON = 3
DISCOUNT = 1
MAX_STEPS = 3
if __name__ == '__main__':
# create world and add agent
world = World()
agent = Agent('Agent')
world.addAgent(agent)
# set parameters
agent.setAttribute('discount', DISCOUNT)
agent.setHorizon(HORIZON)
# add position variable
pos = world.defineState(agent.name, 'position', int, lo=-100, hi=100)
world.setFeature(pos, 0)
# define agents' actions (stay 0, left -1 and right +1)
action = agent.addAction({'verb': 'move', 'action': 'nowhere'})
tree = makeTree(setToFeatureMatrix(pos, pos))
world.setDynamics(pos, action, tree)
action = agent.addAction({'verb': 'move', 'action': 'left'})
tree = makeTree(incrementMatrix(pos, -1))
world.setDynamics(pos, action, tree)
action = agent.addAction({'verb': 'move', 'action': 'right'})
tree = makeTree(incrementMatrix(pos, 1))
def createWorld(username='******',level=0,ability='good',explanation='none',
embodiment='robot',acknowledgment='no',sequence=False,
root='.',ext='xml',beliefs=True):
"""
Creates the initial PsychSim scenario and saves it
@param username: name of user ID to use in filenames
@param level: robot mission level to use as template
@param ability: the level of the robot's ability
- good or C{True}: perfect sensors and sensor model
- badSensor or C{False}: noisy sensors, but perfect model of noisy sensors
- badModel: perfect sensors, but imperfect model of sensors
@type ability: bool
@param explanation: the type of explanation to use
- none: No explanations
- ability: Explanation based on robot ability provided.
- abilitybenevolence: Explanation based on both robot's ability and benevolence provided.
@type explanation: str
@param embodiment: the robot's embodiment
- robot: The robot looks like a robot
- dog: The robot looks like a dog
@type embodiment: str
@param acknowledgment: the robot's behavior regarding the acknowledgment of errors
- no: The robot does not acknowledge its errors
- yes: The robot acknowledges its errors
@type acknowledgment: str
@param root: the root directory to use for files (default is current working directory)
@param ext: the file extension for the PsychSim scenario file
- xml: Save as uncompressed XML
- psy: Save as bzipped XML
@type ext: str
@param beliefs: if C{True}, store robot's uncertain beliefs in scenario file, rather than compute them on the fly. Storing in scenario file makes the scenario a more complete model, but greatly increases file sixe. Default is C{True}
@type beliefs: bool
"""
print "**************************createWorld***********************"
print 'Username:\t%s\nLevel:\t\t%s' % (username,level+1)
print 'Ability\t\t%s\nExplanation:\t%s\nEmbodiment:\t%s\nAcknowledge:\t%s' % \
(ability,explanation,embodiment,acknowledgment)
# Pre-compute symbols for this level's waypoints
for point in WAYPOINTS[level]:
if not point.has_key('symbol'):
point['symbol'] = point['name'].replace(' ','')
world = World()
world.defineState(None,'level',int,lo=0,hi=len(WAYPOINTS)-1,
description='Static variable indicating what mission level')
world.setState(None,'level',level)
world.defineState(None,'time',float)
world.setState(None,'time',0.)
world.defineState(None,'complete',bool)
world.setState(None,'complete',False)
world.addTermination(makeTree({'if': trueRow('complete'), True: True, False: False}))
# Buildings
threats = ['none','NBC','armed']
for waypoint in WAYPOINTS[level]:
if not waypoint.has_key('symbol'):
waypoint['symbol'] = waypoint['name'].replace(' ','')
world.addAgent(Agent(waypoint['symbol']))
# Have we visited this waypoint?
key = world.defineState(waypoint['symbol'],'visited',bool)
world.setFeature(key,False)
# Are there dangerous chemicals or armed people here?
key = world.defineState(waypoint['symbol'],'danger',list,threats[:])
if waypoint.has_key('NBC') and waypoint['NBC']:
world.setFeature(key,'NBC')
elif waypoint.has_key('armed') and waypoint['armed']:
world.setFeature(key,'armed')
else:
world.setFeature(key,'none')
key = world.defineState(waypoint['symbol'],'recommendation',list,
['none','protected','unprotected'])
world.setFeature(key,'none')
# Human
human = Agent('human')
world.addAgent(human)
world.defineState(human.name,'alive',bool)
human.setState('alive',True)
world.defineState(human.name,'deaths',int)
human.setState('deaths',0)
# Robot
robot = Agent('robot')
world.addAgent(robot)
# Robot states
world.defineState(robot.name,'waypoint',list,[point['symbol'] for point in WAYPOINTS[level]])
robot.setState('waypoint',WAYPOINTS[level][getStart(level)]['symbol'])
world.defineState(robot.name,'explanation',list,['none','ability','abilitybenevolence','abilityconfidence','confidence'])
robot.setState('explanation',explanation)
world.defineState(robot.name,'embodiment',list,['robot','dog'])
robot.setState('embodiment',embodiment)
world.defineState(robot.name,'acknowledgment',list,['no','yes'])
robot.setState('acknowledgment',acknowledgment)
world.defineState(robot.name,'ability',list,['badSensor','badModel','good'])
if ability is True:
# Backward compatibility with boolean ability
ability = 'good'
elif ability is False:
ability = 'badSensor'
robot.setState('ability',ability)
# State of the robot's sensors
world.defineState(robot.name,'sensorModel',list,['good','bad'])
robot.setState('sensorModel','good')
world.defineState(robot.name,'command',list,['none']+[point['symbol'] for point in WAYPOINTS[level]])
robot.setState('command','none')
# Actions
for end in range(len(WAYPOINTS[level])):
symbol = WAYPOINTS[level][end]['symbol']
# Robot movement
action = robot.addAction({'verb': 'moveto','object': symbol})
# Legal if no contradictory command
tree = makeTree({'if': equalRow(stateKey(robot.name,'command'),'none'),
True: True,
False: {'if': equalRow(stateKey(robot.name,'command'),symbol),
True: True, False: False}})
robot.setLegal(action,tree)
# Dynamics of robot's location
tree = makeTree(setToConstantMatrix(stateKey(action['subject'],'waypoint'),symbol))
world.setDynamics(stateKey(action['subject'],'waypoint'),action,tree)
# Dynamics of visited flag
key = stateKey(symbol,'visited')
tree = makeTree(setTrueMatrix(key))
world.setDynamics(key,action,tree)
# Dynamics of time
key = stateKey(None,'time')
tree = setToConstantMatrix(key,0.)
for start in range(len(WAYPOINTS[level])):
if start != end:
startsymbol = WAYPOINTS[level][start]['symbol']
if sequence:
# Distance is measured by level sequence
distance = abs(start-end)*50
else:
try:
distance = DISTANCES[WAYPOINTS[level][start]['name']][WAYPOINTS[level][end]['name']]
except KeyError:
try:
distance = DISTANCES[WAYPOINTS[level][end]['name']][WAYPOINTS[level][start]['name']]
except KeyError:
distance = 250
tree = {'if': equalRow(stateKey(action['subject'],'waypoint'),startsymbol),
True: setToConstantMatrix(key,float(distance)/1000.),
False: tree}
world.setDynamics(key,action,makeTree(tree))
# Human entry: Dead or alive if unprotected?
key = stateKey(human.name,'alive')
action = robot.addAction({'verb': 'recommend unprotected','object': symbol})
tree = makeTree({'if': equalRow(stateKey(symbol,'danger'),'none'),
True: setTrueMatrix(key), False: setFalseMatrix(key)})
world.setDynamics(key,action,tree)
robot.setLegal(action,makeTree(False))
# Human entry: How much "time" if protected?
action = robot.addAction({'verb': 'recommend protected','object': symbol})
key = stateKey(None,'time')
world.setDynamics(key,action,makeTree(setToConstantMatrix(key,0.25)))
robot.setLegal(action,makeTree(False))
# Robot goals
goal = minimizeFeature(stateKey(None,'time'))
robot.setReward(goal,2.)
goal = maximizeFeature(stateKey(human.name,'alive'))
robot.setReward(goal,1.)
for point in WAYPOINTS[level]:
robot.setReward(maximizeFeature(stateKey(point['symbol'],'visited')),1.)
if beliefs:
# omega = 'danger'
world.defineVariable(robot.name,ActionSet)
# Robot beliefs
world.setModel(robot.name,True)
value = 1./float(len(WAYPOINTS[level]))
# tree = KeyedVector({CONSTANT: world.value2float(omega,'none')})
for index in range(len(WAYPOINTS[level])):
waypoint = WAYPOINTS[level][index]
key = stateKey(waypoint['symbol'],'danger')
# if index > 0:
# Starting state is safe
robot.setBelief(key,psychsim.probability.Distribution({'NBC': value/2., 'armed': value/2.,'none': 1.-value}))
# Observation function
# tree = {'if': equalRow(stateKey(robot.name,'waypoint'),waypoint['symbol']),
# True: generateO(world,key),
# False: tree}
# robot.defineObservation(omega,makeTree(tree),domain=list,lo=['none','NBC','armed'])
robot.defineObservation('microphone',makeTree(None),None,domain=list,
lo=['nobody','friendly','suspicious'])
robot.defineObservation('NBCsensor',makeTree(None),None,domain=bool)
robot.defineObservation('camera',makeTree(None),None,domain=bool)
else:
robot.defineObservation('microphone',makeTree(None),None,domain=list,
lo=['nobody','friendly','suspicious'])
robot.defineObservation('NBCsensor',makeTree(None),None,domain=bool)
robot.defineObservation('camera',makeTree(None),None,domain=bool)
robot.setAttribute('horizon',1)
world.setOrder([robot.name])
filename = getFilename(username,level,ext,root)
world.save(filename,ext=='psy')
WriteLogData('%s user %s, level %d, ability %s, explanation %s' % \
(CREATE_TAG,username,level,ability,explanation),username,level,root=root)
return world
world.setDynamics(location, behaviors['leave']['action'],
makeTree(setToConstantMatrix(location, 'beyond')))
world.setDynamics(location, behaviors['return']['action'],
makeTree(setToConstantMatrix(location, 'Seattle')))
# Phase dynamics
world.setDynamics(
'phase', True,
makeTree({
'if': equalRow('phase', 'where'),
True: setToConstantMatrix('phase', 'how'),
False: setToConstantMatrix('phase', 'where')
}))
# Decision-making parameters
resident.setAttribute('horizon', 2)
resident.setAttribute('selection', 'distribution')
world.save('anthrax.psy')
# world.printState()
for tree, weight in list(resident.getAttribute('R').items()):
print(weight, tree)
decision = Distribution()
for vector in world.state[None].domain():
world.printVector(vector)
result = resident.decide(vector, selection='distribution')
for action in result['action'].domain():
decision.addProb(
action, world.state[None][vector] * result['action'][action])
def setup():
global args
np.random.seed(args.seed)
# create world and add agents
world = World()
world.memory = False
world.parallel = args.parallel
agents = []
agent_features = {}
for ag in range(args.agents):
agent = Agent('Agent' + str(ag))
world.addAgent(agent)
agents.append(agent)
# set agent's params
agent.setAttribute('discount', 1)
agent.setHorizon(args.horizon)
# add features, initialize at random
features = []
agent_features[agent] = features
for f in range(args.features_agent):
feat = world.defineState(agent.name, 'Feature{}'.format(f), int, lo=0, hi=1000)
world.setFeature(feat, np.random.randint(0, MAX_FEATURE_VALUE))
features.append(feat)
# set random reward function
agent.setReward(maximizeFeature(np.random.choice(features), agent.name), 1)
# add mental copy of true model and make it static (we do not have beliefs in the models)
agent.addModel(get_fake_model_name(agent), parent=get_true_model_name(agent))
agent.setAttribute('static', True, get_fake_model_name(agent))
# add actions
for ac in range(args.actions):
action = agent.addAction({'verb': '', 'action': 'Action{}'.format(ac)})
i = ac
while i + args.features_action < args.features_agent:
weights = {}
for j in range(args.features_action):
weights[features[i + j + 1]] = 1
tree = makeTree(multi_set_matrix(features[i], weights))
world.setDynamics(features[i], action, tree)
i += args.features_action
# define order
world.setOrder([set(ag.name for ag in agents)])
for agent in agents:
# test belief update:
# - set a belief in one feature to the actual initial value (should not change outcomes)
# world.setModel(agent.name, Distribution({True: 1.0}))
rand_feat = np.random.choice(agent_features[agent])
agent.setBelief(rand_feat, world.getValue(rand_feat))
print('{} will always observe {}={}'.format(agent.name, rand_feat, world.getValue(rand_feat)))
# set mental model of each agent in all other agents
for i in range(args.agents):
for j in range(i + 1, args.agents):
world.setMentalModel(agents[i].name, agents[j].name, Distribution({get_fake_model_name(agents[j]): 1}))
world.setMentalModel(agents[j].name, agents[i].name, Distribution({get_fake_model_name(agents[i]): 1}))
return world
class TestAgents(unittest.TestCase):
def setUp(self):
# Create world
self.world = World()
# Create agents
self.tom = Agent('Tom')
self.world.addAgent(self.tom)
self.jerry = Agent('Jerry')
self.world.addAgent(self.jerry)
def addStates(self):
"""Create state features"""
self.world.defineState(self.tom.name,'health',int,lo=0,hi=100,
description='%s\'s wellbeing' % (self.tom.name))
self.world.setState(self.tom.name,'health',50)
self.world.defineState(self.jerry.name,'health',int,lo=0,hi=100,
description='%s\'s wellbeing' % (self.jerry.name))
self.world.setState(self.jerry.name,'health',50)
def addActions(self):
"""Create actions"""
self.chase = self.tom.addAction({'verb': 'chase','object': self.jerry.name})
self.hit = self.tom.addAction({'verb': 'hit','object': self.jerry.name})
self.run = self.jerry.addAction({'verb': 'run away'})
self.trick = self.jerry.addAction({'verb': 'trick','object': self.tom.name})
def addDynamics(self):
"""Create dynamics"""
tree = makeTree(incrementMatrix(stateKey(self.jerry.name,'health'),-10))
self.world.setDynamics(stateKey(self.jerry.name,'health'),self.hit,tree,enforceMin=True)
def addModels(self,rationality=1.):
self.tom.addModel('friend',rationality=rationality,parent=True)
self.tom.setReward(maximizeFeature(stateKey(self.jerry.name,'health')),1.,'friend')
self.tom.addModel('foe',rationality=rationality,parent=True)
self.tom.setReward(minimizeFeature(stateKey(self.jerry.name,'health')),1.,'foe')
def saveload(self):
"""Write scenario to file and then load from scratch"""
self.world.save('/tmp/psychsim_test.psy')
self.world = World('/tmp/psychsim_test.psy')
self.tom = self.world.agents[self.tom.name]
self.jerry = self.world.agents[self.jerry.name]
def testEnumeratedState(self):
self.addActions()
self.world.defineVariable(self.tom.name,ActionSet)
self.world.defineState(self.tom.name,'status',list,['dead','injured','healthy'])
self.world.setState(self.tom.name,'status','healthy')
goal = achieveFeatureValue(stateKey(self.tom.name,'status'),'healthy')
self.tom.setReward(goal,1.)
goal = achieveFeatureValue(stateKey(self.tom.name,'status'),'injured')
self.jerry.setReward(goal,1.)
self.saveload()
self.assertEqual(len(self.world.state),1)
vector = self.world.state.domain()[0]
tVal = self.tom.reward(vector)
self.assertAlmostEqual(tVal,1.,8)
jVal = self.jerry.reward(vector)
self.assertAlmostEqual(jVal,0.,8)
for action in self.tom.actions:
encoding = self.world.value2float(self.tom.name,action)
self.assertEqual(action,self.world.float2value(self.tom.name,encoding))
def testBeliefModels(self):
self.addStates()
self.addActions()
self.addDynamics()
self.world.setOrder([self.tom.name])
self.tom.addModel('optimist')
self.tom.setBelief(stateKey(self.jerry.name,'health'),20,'optimist')
self.tom.addModel('pessimist')
self.world.setModel(self.jerry.name,True)
self.world.setMentalModel(self.jerry.name,self.tom.name,{'optimist': 0.5,'pessimist': 0.5})
actions = {self.tom.name: self.hit}
self.world.step(actions)
vector = self.world.state.domain()[0]
beliefs = self.jerry.getAttribute('beliefs',self.world.getModel(self.jerry.name,vector))
for belief in beliefs.domain():
model = self.world.getModel(self.tom.name,belief)
if self.tom.models[model].has_key('beliefs'):
nested = self.tom.models[model]['beliefs']
self.assertEqual(len(nested),1)
nested = nested.domain()[0]
self.assertEqual(len(nested),1)
self.assertAlmostEqual(nested[stateKey(self.jerry.name,'health')],10.,8)
def testObservation(self):
self.addStates()
self.addActions()
self.addDynamics()
self.world.setOrder([self.tom.name])
self.world.setModel(self.jerry.name,True)
key = stateKey(self.jerry.name,'health')
self.jerry.setBelief(key,Distribution({20: 0.5, 50: 0.5}))
tree = makeTree({'if': thresholdRow(key,40),
True: {'distribution': [(KeyedVector({CONSTANT: 50}),.8),
(KeyedVector({CONSTANT: 20}),.2)]},
False: {'distribution': [(KeyedVector({CONSTANT: 50}),.2),
(KeyedVector({CONSTANT: 20}),.8)]}})
self.jerry.defineObservation(key,tree)
actions = {self.tom.name: self.hit}
vector = self.world.state.domain()[0]
omegaDist = self.jerry.observe(vector,actions)
for omega in omegaDist.domain():
new = KeyedVector(vector)
model = self.jerry.index2model(self.jerry.stateEstimator(vector,new,omega))
beliefs = self.jerry.models[model]['beliefs']
if omega[key] > 30:
# We observed a high value, so we should have a stronger belief in the higher value
# which is now 40 after the hit
for belief in beliefs.domain():
if beliefs[belief] > 0.5:
self.assertAlmostEqual(belief[key],40,8)
else:
self.assertAlmostEqual(belief[key],10,8)
else:
# We observed a low value, so we should have a stronger belief in the lower value
# which is now 10 after the hit
for belief in beliefs.domain():
if beliefs[belief] < 0.5:
self.assertAlmostEqual(belief[key],40,8)
else:
self.assertAlmostEqual(belief[key],10,8)
def testUnobservedAction(self):
self.addStates()
self.addActions()
self.addDynamics()
self.addModels()
self.world.setOrder([self.tom.name])
self.world.setModel(self.jerry.name,True)
self.jerry.setBelief(stateKey(self.jerry.name,'health'),50)
self.world.setMentalModel(self.jerry.name,self.tom.name,{'friend': 0.5,'foe': 0.5})
tree = makeTree(True)
self.jerry.defineObservation(self.tom.name,tree,self.hit,domain=ActionSet)
tree = makeTree({'distribution': [(True,0.25),(False,0.75)]})
self.jerry.defineObservation(self.tom.name,tree,self.chase,domain=ActionSet)
vector = self.world.state.domain()[0]
self.saveload()
self.world.step({self.tom.name: self.hit})
vector = self.world.state.domain()[0]
def testRewardModels(self):
self.addStates()
self.addActions()
self.addDynamics()
self.addModels()
self.world.setOrder([self.tom.name])
# Add Jerry's model to the world (so that it gets updated)
self.world.setModel(self.jerry.name,True)
# Give Jerry uncertainty about Tom
self.world.setMentalModel(self.jerry.name,self.tom.name,{'friend': 0.5,'foe': 0.5})
self.saveload()
# Hitting should make Jerry think Tom is more of a foe
actions = {self.tom.name: self.hit}
self.world.step(actions)
vector = self.world.state.domain()[0]
belief01 = self.jerry.getAttribute('beliefs',self.world.getModel(self.jerry.name,vector))
key = modelKey(self.tom.name)
for belief in belief01.domain():
if self.tom.index2model(belief[key]) == 'foe':
prob01 = belief01[belief]
break
self.assertGreater(prob01,0.5)
# If we think of Tom as even more of an optimizer, then our update should be stronger
self.tom.setAttribute('rationality',10.,'foe')
self.tom.setAttribute('rationality',10.,'friend')
self.world.setMentalModel(self.jerry.name,self.tom.name,{'friend': 0.5,'foe': 0.5})
self.world.step(actions)
vector = self.world.state.domain()[0]
model = self.world.getModel(self.jerry.name,vector)
belief10 = self.jerry.getAttribute('beliefs',model)
key = modelKey(self.tom.name)
for belief in belief10.domain():
if self.tom.index2model(belief[key]) == 'foe':
prob10 = belief10[belief]
break
self.assertGreater(prob10,prob01)
# If we keep the same models, but get another observation, we should update even more
self.world.step(actions)
vector = self.world.state.domain()[0]
model = self.world.getModel(self.jerry.name,vector)
belief1010 = self.jerry.getAttribute('beliefs',model)
key = modelKey(self.tom.name)
for belief in belief1010.domain():
if self.tom.index2model(belief[key]) == 'foe':
prob1010 = belief1010[belief]
break
self.assertGreater(prob1010,prob10)
def testDynamics(self):
self.world.setOrder([self.tom.name])
self.addStates()
self.addActions()
self.addDynamics()
key = stateKey(self.jerry.name,'health')
self.assertEqual(len(self.world.state),1)
vector = self.world.state.domain()[0]
self.assertTrue(vector.has_key(stateKey(self.tom.name,'health')))
self.assertTrue(vector.has_key(turnKey(self.tom.name)))
self.assertTrue(vector.has_key(key))
self.assertTrue(vector.has_key(CONSTANT))
self.assertEqual(len(vector),4)
self.assertEqual(vector[stateKey(self.tom.name,'health')],50)
self.assertEqual(vector[key],50)
outcome = self.world.step({self.tom.name: self.chase})
for i in range(7):
self.assertEqual(len(self.world.state),1)
vector = self.world.state.domain()[0]
self.assertTrue(vector.has_key(stateKey(self.tom.name,'health')))
self.assertTrue(vector.has_key(turnKey(self.tom.name)))
self.assertTrue(vector.has_key(key))
self.assertTrue(vector.has_key(CONSTANT))
self.assertEqual(len(vector),4)
self.assertEqual(vector[stateKey(self.tom.name,'health')],50)
self.assertEqual(vector[key],max(50-10*i,0))
outcome = self.world.step({self.tom.name: self.hit})
self.saveload()
def testRewardOnOthers(self):
self.addStates()
self.addActions()
self.addDynamics()
self.world.setOrder([self.tom.name])
vector = self.world.state.domain()[0]
# Create Jerry's goals
goal = maximizeFeature(stateKey(self.jerry.name,'health'))
self.jerry.setReward(goal,1.)
jVal = -self.jerry.reward(vector)
# Create Tom's goals from scratch
minGoal = minimizeFeature(stateKey(self.jerry.name,'health'))
self.tom.setReward(minGoal,1.)
self.saveload()
tRawVal = self.tom.reward(vector)
self.assertAlmostEqual(jVal,tRawVal,8)
# Create Tom's goals as a function of Jerry's
self.tom.models[True]['R'].clear()
self.tom.setReward(self.jerry.name,-1.)
self.saveload()
tFuncVal = self.tom.reward(vector)
self.assertAlmostEqual(tRawVal,tFuncVal,8)
# Test effect of functional reward on value function
self.tom.setHorizon(1)
self.saveload()
vHit = self.tom.value(vector,self.hit)['V']
vChase = self.tom.value(vector,self.chase)['V']
self.assertAlmostEqual(vHit,vChase+.1,8)
def testReward(self):
self.addStates()
key = stateKey(self.jerry.name,'health')
goal = makeTree({'if': thresholdRow(key,5),
True: KeyedVector({key: -2}),
False: KeyedVector({key: -1})})
self.jerry.setReward(goal,1.)
R = self.jerry.models[True]['R']
self.assertEqual(len(R),1)
self.assertEqual(R.keys()[0],goal)
self.assertAlmostEqual(R[goal],1.,8)
self.jerry.setReward(goal,2.)
self.assertEqual(len(R),1)
self.assertEqual(R.keys()[0],goal)
self.assertAlmostEqual(R[goal],2.,8)
def testTurnDynamics(self):
self.addStates()
self.addActions()
self.world.setOrder([self.tom.name,self.jerry.name])
self.assertEqual(self.world.maxTurn,1)
self.saveload()
vector = self.world.state.domain()[0]
jTurn = turnKey(self.jerry.name)
tTurn = turnKey(self.tom.name)
self.assertEqual(self.world.next(),[self.tom.name])
self.assertEqual(vector[tTurn],0)
self.assertEqual(vector[jTurn],1)
self.world.step()
vector = self.world.state.domain()[0]
self.assertEqual(self.world.next(),[self.jerry.name])
self.assertEqual(vector[tTurn],1)
self.assertEqual(vector[jTurn],0)
self.world.step()
vector = self.world.state.domain()[0]
self.assertEqual(self.world.next(),[self.tom.name])
self.assertEqual(vector[tTurn],0)
self.assertEqual(vector[jTurn],1)
# Try some custom dynamics
self.world.setTurnDynamics(self.tom.name,self.hit,makeTree(noChangeMatrix(tTurn)))
self.world.setTurnDynamics(self.jerry.name,self.hit,makeTree(noChangeMatrix(tTurn)))
self.world.step()
vector = self.world.state.domain()[0]
self.assertEqual(self.world.next(),[self.tom.name])
self.assertEqual(vector[tTurn],0)
self.assertEqual(vector[jTurn],1)
self.world.step({self.tom.name: self.chase})
vector = self.world.state.domain()[0]
self.assertEqual(self.world.next(),[self.jerry.name])
self.assertEqual(vector[tTurn],1)
self.assertEqual(vector[jTurn],0)
def testStatic(self):
self.addStates()
self.addActions()
self.addDynamics()
self.addModels()
self.world.setModel(self.jerry.name,True)
self.world.setMentalModel(self.jerry.name,self.tom.name,{'friend': 0.5,'foe': 0.5})
self.world.setOrder([self.tom.name])
vector = self.world.state.domain()[0]
model = self.world.getModel(self.jerry.name,vector)
belief0 = self.jerry.models[model]['beliefs']
self.world.step()
vector = self.world.state.domain()[0]
model = self.world.getModel(self.jerry.name,vector)
belief1 = self.jerry.models[model]['beliefs']
key = modelKey(self.tom.name)
for vector in belief0.domain():
if self.tom.index2model(vector[key]) == 'friend':
self.assertGreater(belief0[vector],belief1[vector])
else:
self.assertGreater(belief1[vector],belief0[vector])
# Now with the static beliefs
self.jerry.setAttribute('static',True,model)
self.saveload()
self.world.step()
vector = self.world.state.domain()[0]
model = self.world.getModel(self.jerry.name,vector)
belief2 = self.jerry.models[model]['beliefs']
for vector in belief1.domain():
self.assertAlmostEqual(belief1[vector],belief2[vector],8)