def __init__(self,name,imagefile,context,skills=[],factionid=0):
"""
"""
model = [ (0,0), (0,181), (100,181), (100,0) ]
bounding_rect = [ (0,0), (0,181), (100,181), (100,0) ]
#super( Character, self ).__init__( model, imagefile, offset=(0,50) )
_model = Vertel()
_model.set( [ ( 0.0, 0.0, 0.0 ),
( 0.0, 171.0/600.0, 0.0 ),
( 101.0/800.0, 171.0/600.0,0.0 ),
( 101.0/800.0 , 0.0, 0.0 ) ] )
_bound = Vertel()
_bound.set( [ ( 0.0, 0.0, 0.0 ),
( 0.0, 100.0/600, 0.0 ),
( 100.0/800.0, 100.0/600, 0.0 ),
( 100.0/800, 0.0, 0.0 ) ] )
super( Character, self ).__init__( imagefile, _model, _bound )
#self.name = kwargs.get('name',"Unnamed")
self.name = name
self.mind = Mind( ego=self, factionid=factionid )
self._controller = None
self._context = context
self._target = None
#body stats
self.strength = 10 #strength of character, in kg*m*s^-2
self.agility = 20 #agility of character, in kg*m*s^-2
self.endurance = 10 #endurance: MAX Energy Reserves, in
#endurance
#endurance: MAX energy output in watts in kg*m^2*s^-3
self._reach = 1.0 #arm length
self._mass = 50 #mass of character, in kg
self.health = 100 #
self.energy = Energy()
self._alive = True #alive bit, used to drop a character object references
#skills
self.skillbook = SkillBook(skills=skills)
self.deck = Deck(maxsize=4, slots=['default']*6 )
#equipment
self._equipment = Equipment() #equipment is currently held by the character
self.inventory = Inventory(40.0) #items in storage
self.weapon = None
#modifiers
self.buff = Buffable(self) #allows the character to be buffed (stat and status effects)
self.status_effects = []
self._wounds = []
def __init__(self, envName, numThreads=3):
self.lock_queue = threading.Lock()
self.enviromentName = envName
self.env = Enviroment(envName)
self.mind = Mind(self.env.envDim(),
self.env.actDim(),
policyOptimizer=Adam(lr=0.001),
criticOptimizer=Adam(lr=0.001))
self.optimizer = A3C.OptimizerThread(self.mind)
self.numThreads = numThreads
self.results = History()
commons.tensorFlowGraph = K.get_session().graph
def main():
m = Mind()
clauses = [
DeepClause('see', {
'agent': 'Tim',
'target': 'Tom',
}),
DeepClause('see', {
'agent': UNK,
'target': 'Tom',
}),
]
for c in clauses:
print '-' * 80
print c.to_d()
for clause in m.receive(c):
print '*', json.dumps(clause.to_d(), indent=4)
class Agent:
def __init__(self):
self.mind = Mind(self)
self.init_agent()
def init_agent(self):
self.drawer = Drawer(self)
self.goal = 0
self.goal_integral = 0
self.foods_eaten = 0
self.food_pos = [random.uniform(0, 1) for _ in range(2)]
self.agent_pos = [random.uniform(0, 1) for _ in range(2)]
self.agent_vel = [random.uniform(0, 1) for _ in range(2)]
def live(self, generation):
for i in range(iterations):
self.mind.one_iteration(generation)
return self.goal_integral
def percept(self):
distance_to_food = [(self.food_pos[i] - self.agent_pos[i])
for i in range(2)]
self.goal = (self.agent_pos[0] - self.agent_pos[1]
) / 2 #sum([pos for pos in self.agent_pos]) / 2
#return [self.goal] + distance_to_food + self.agent_vel
return [self.goal] + self.agent_vel
def update_sensors(self):
#self.goal -= goal_decay if self.goal == 0 else 0
self.goal_integral += self.goal
for i in range(2):
self.agent_pos[i] += self.agent_vel[i] / 20
if not (0 < self.agent_pos[i] < 1):
self.agent_pos[i] = int(
self.agent_pos[i]) #.5 + random.uniform(-.1, .1)#
'''
class Prelude(object):
def __init__(self):
self.timer = None
self.isSpeaking = False
self.proactiveMode = False
self.brainLocation = "mind.mdu"
self.quantumRandomness = False
self.avoidLearnByRepeating = False
self.associater = MatchingAlgorithm.CosineTFIDF
self.saveTracking = time.time()
def initializeEngine(self):
self.mindInstance = Mind()
self.mindInstance.associater = self.associater
self.mindInstance.brainLocation = self.brainLocation
self.chatInitiated = time.strftime("%H:%M:%S")
self.mindInstance.analyzeShortTermMemory()
def chatWithPrelude(self, question):
if self.mindInstance is None:
return "Error: Mind not initialized";
if self.proactiveMode:
self.idleTime = 0
if self.timer is not None:
timer.Stop()
if self.quantumRandomness:
self.mindInstance.quantumRandomness = True
if self.avoidLearnByRepeating:
self.mindInstance.avoidLearnByRepeating = True
answer = ""
answer = self.mindInstance.listenToInput(question)
if self.isSpeaking:
self.speak(answer)
if self.proactiveMode:
self.setTimer()
self.autoSpeakInput = answer
self.timer.start()
elapsed_time = time.time() - self.saveTracking
if elapsed_time > 10:
print "soft save"
self.forceSaveMindFile()
self.saveTracking = time.time()
return answer
def setTimer(self):
pass
def monoLog(self):
pass
def speak(self, a):
pass
def stopEngine(self):
if self.mindInstance:
self.mindInstance.prepareCurrentMemoryForDisc()
if self.isContributable:
self.mindInstance.contributeClientMind()
def forceUpload(self):
pass
def forceSaveMindFile(self):
if self.mindInstance:
self.mindInstance.prepareCurrentMemoryForDisc()
def countMindMemory(self):
if self.mindInstance:
print "My memory contains " + str(len(self.mindInstance.botsMemory)) + " neurons"
return str(len(self.mindInstance.botsMemory))
def setTimer(self):
self.timer = Timer(20 * 60, autoAnswer)
def autoAnswer(self):
try:
t = self.mindInstance.listenToInput(self.autoSpeakInput)
print "You (away):\t"
print "Prelude (bored):\t" + t
except:
print "Error: autoAnswer"
def setAssociationAlgorithm(self, assocType):
# if self.mindInstance:
# self.mindInstance.associater = assocType
self.associater = assocType
def setMindFileLocation(self, fileName):
self.brainLocation = fileName
def getVersionInfo(self):
return "Prelude@# Engine, version 1.2.7, 2004-2015(c) by Lennart Lopin ";
def setBotClient(self):
pass
def setProactiveMode(self, flag):
pass
def initializeEngine(self):
self.mindInstance = Mind()
self.mindInstance.associater = self.associater
self.mindInstance.brainLocation = self.brainLocation
self.chatInitiated = time.strftime("%H:%M:%S")
self.mindInstance.analyzeShortTermMemory()
class Prelude(object):
def __init__(self):
self.timer = None
self.isSpeaking = False
self.proactiveMode = False
self.brainLocation = "mind.mdu"
self.quantumRandomness = False
self.avoidLearnByRepeating = False
self.associater = MatchingAlgorithm.CosineTFIDF
self.saveTracking = time.time()
def initializeEngine(self):
self.mindInstance = Mind()
self.mindInstance.associater = self.associater
self.mindInstance.brainLocation = self.brainLocation
self.chatInitiated = time.strftime("%H:%M:%S")
self.mindInstance.analyzeShortTermMemory()
def chatWithPrelude(self, question):
if self.mindInstance is None:
return "Error: Mind not initialized"
if self.proactiveMode:
self.idleTime = 0
if self.timer is not None:
timer.Stop()
if self.quantumRandomness:
self.mindInstance.quantumRandomness = True
if self.avoidLearnByRepeating:
self.mindInstance.avoidLearnByRepeating = True
answer = ""
answer = self.mindInstance.listenToInput(question)
if self.isSpeaking:
self.speak(answer)
if self.proactiveMode:
self.setTimer()
self.autoSpeakInput = answer
self.timer.start()
elapsed_time = time.time() - self.saveTracking
if elapsed_time > 10:
print "soft save"
self.forceSaveMindFile()
self.saveTracking = time.time()
return answer
def setTimer(self):
pass
def monoLog(self):
pass
def speak(self, a):
pass
def stopEngine(self):
if self.mindInstance:
self.mindInstance.prepareCurrentMemoryForDisc()
if self.isContributable:
self.mindInstance.contributeClientMind()
def forceUpload(self):
pass
def forceSaveMindFile(self):
if self.mindInstance:
self.mindInstance.prepareCurrentMemoryForDisc()
def countMindMemory(self):
if self.mindInstance:
print "My memory contains " + str(len(
self.mindInstance.botsMemory)) + " neurons"
def setTimer(self):
self.timer = Timer(20 * 60, autoAnswer)
def autoAnswer(self):
try:
t = self.mindInstance.listenToInput(self.autoSpeakInput)
print "You (away):\t"
print "Prelude (bored):\t" + t
except:
print "Error: autoAnswer"
def setAssociationAlgorithm(self, assocType):
# if self.mindInstance:
# self.mindInstance.associater = assocType
self.associater = assocType
def setMindFileLocation(self, fileName):
self.brainLocation = fileName
def getVersionInfo(self):
return "Prelude@# Engine, version 1.2.7, 2004-2015(c) by Lennart Lopin "
def setBotClient(self):
pass
def setProactiveMode(self, flag):
pass
def initializeEngine(self):
self.mindInstance = Mind()
self.mindInstance.associater = self.associater
self.mindInstance.brainLocation = self.brainLocation
self.chatInitiated = time.strftime("%H:%M:%S")
self.mindInstance.analyzeShortTermMemory()
class Character( Actor ):
"""
"""
#===============================================================================================
def __init__(self,name,imagefile,context,skills=[],factionid=0):
"""
"""
model = [ (0,0), (0,181), (100,181), (100,0) ]
bounding_rect = [ (0,0), (0,181), (100,181), (100,0) ]
#super( Character, self ).__init__( model, imagefile, offset=(0,50) )
_model = Vertel()
_model.set( [ ( 0.0, 0.0, 0.0 ),
( 0.0, 171.0/600.0, 0.0 ),
( 101.0/800.0, 171.0/600.0,0.0 ),
( 101.0/800.0 , 0.0, 0.0 ) ] )
_bound = Vertel()
_bound.set( [ ( 0.0, 0.0, 0.0 ),
( 0.0, 100.0/600, 0.0 ),
( 100.0/800.0, 100.0/600, 0.0 ),
( 100.0/800, 0.0, 0.0 ) ] )
super( Character, self ).__init__( imagefile, _model, _bound )
#self.name = kwargs.get('name',"Unnamed")
self.name = name
self.mind = Mind( ego=self, factionid=factionid )
self._controller = None
self._context = context
self._target = None
#body stats
self.strength = 10 #strength of character, in kg*m*s^-2
self.agility = 20 #agility of character, in kg*m*s^-2
self.endurance = 10 #endurance: MAX Energy Reserves, in
#endurance
#endurance: MAX energy output in watts in kg*m^2*s^-3
self._reach = 1.0 #arm length
self._mass = 50 #mass of character, in kg
self.health = 100 #
self.energy = Energy()
self._alive = True #alive bit, used to drop a character object references
#skills
self.skillbook = SkillBook(skills=skills)
self.deck = Deck(maxsize=4, slots=['default']*6 )
#equipment
self._equipment = Equipment() #equipment is currently held by the character
self.inventory = Inventory(40.0) #items in storage
self.weapon = None
#modifiers
self.buff = Buffable(self) #allows the character to be buffed (stat and status effects)
self.status_effects = []
self._wounds = []
#===============================================================================================
# Helper functions which link related members of a Character object, such as the skll book and
# skill deck
#===============================================================================================
def setslot(self,**kwargs):
"""
helper function, sets deck slot type
"""
self.deck.setslot(**kwargs)
def setskill(self,**kwargs):
"""
helper function, sets deck slot with skill from skillbook
"""
self.deck.setskill(slot=kwargs['slot'],skill=self.skillbook[kwargs['skill']])
def queueskill(self,**kwargs):
"""
helper function, queues a skill as the next action from the deck
"""
print "Character queue skill"
try:
skill = self.deck[kwargs['slot']]
if skill:
self.mind.appendaction( Action( actor = self,
atype = skill.type,
target = self.mind._target,
skill = skill,
stage = self._context,
energy = 3 ) )
else:
print "No Skill set for deck slot:", kwargs['slot']
except KeyError:
print "KeyError in Character:",self.name," queueskill"
print " Either 'slot':'val' not present in arguments, or 'slot' not in deck"
print " Keys in deck", self.deck.keys()
print " kwargs:", kwargs
#===============================================================================================
def reach(self):
"""
return the characters effective reach,
taking into account
- base reach
- weapon reach
"""
return ( self._reach + self._equipment["rhand"]._reach )
#===============================================================================================
def __str__(self):
"""
"""
return 'Character: %(name)s' %{'name':self.name}
#===============================================================================================
def recharge(self,step=1):
"""
recharge is the turn update function, called recharge to avoid ambiguity with the animated
actor's update function. Recharge is called at the beginning of a character's turn.
"""
self.energy.rechargehand() #draw a new hand of energy from the pool
#todo decrement cooldown on abilities by 1
#===============================================================================================
def use(self, energy, attempt=False):
"""
remove the specified amount of energy from this characters' hand
"""
return self.energy.subhand(energy,attempt)
#===============================================================================================
def hand(self):
"""return the current energy left in this characters' hand"""
return self.energy._hand
#===============================================================================================
def pool(self):
"""return the current energy left in this characters' pool"""
return self._energy_pool
#===============================================================================================
def iscapable(self, action):
"""
return True if this character is capable of performing the action, False otherwise
whether a character is capable of performing a given action or not is dependent on the
energy cost of the action, and any relevant status effects on the character
"""
if self.energy._hand >= action.getenergy():
return True
#===============================================================================================
def update(self, step=1):
"""
"""
#call super class update
self.recharge(step)
self.health -= len(self._wounds)*25
#mind.plan
return self.health
#===============================================================================================
def settarget(self, item ):
"""
"""
self._target = item
self.mind._target = item
def gettarget(self):
"""
"""
return self._target
#===============================================================================================
def listen(self):
"""
TODO: Make character an event listener
TODO: this should be an event handler
currently just updates wound and energy
"""
#call super class update
self.recharge(1)
self.health -= len(self._wounds)*25
#mind.plan
if not self.mind.targetalive(): self.mind._target = None
return self.health
class A3C:
class OptimizerThread(threading.Thread):
stop_signal = False
def __init__(self, mind):
self.mind = mind
threading.Thread.__init__(self)
def run(self):
while not self.stop_signal:
self.mind.train()
def stop(self):
self.stop_signal = True
class AgentThread(threading.Thread):
stop_signal = False
def __init__(self, agent, resultHistory):
threading.Thread.__init__(self)
self.agent = agent
self.history = resultHistory
def run(self):
while not self.stop_signal:
result = self.agent.playSingleEpisode()
if result != None:
self.history.registerResult(result)
def stop(self):
self.stop_signal = True
def __init__(self, envName, numThreads=3):
self.lock_queue = threading.Lock()
self.enviromentName = envName
self.env = Enviroment(envName)
self.mind = Mind(self.env.envDim(),
self.env.actDim(),
policyOptimizer=Adam(lr=0.001),
criticOptimizer=Adam(lr=0.001))
self.optimizer = A3C.OptimizerThread(self.mind)
self.numThreads = numThreads
self.results = History()
commons.tensorFlowGraph = K.get_session().graph
def train(self, stopFunction, remapRewardFunction=None, epsilon=1.0):
self.optimizer.start()
agents = list()
for idEnv in range(self.numThreads):
agent = Agent(Enviroment(self.enviromentName, seed=idEnv),
self.mind,
rewardMapper=remapRewardFunction,
epsilon=epsilon,
name="Agent-" + str(idEnv))
agThread = A3C.AgentThread(agent, self.results)
agThread.start()
agents.append(agThread)
done = False
while not done:
time.sleep(1)
done = stopFunction(list(self.results.results))
self.optimizer.stop()
for agent in agents:
agent.stop()
self.mind.save(self.enviromentName)
def play(self):
self.mind.trained = True
self.mind.load_weights(self.enviromentName)
agent = Agent(Enviroment(self.enviromentName), self.mind, epsilon=0.0)
return agent.playSingleEpisode(render=True)
def __init__(self):
self.mind = Mind(self)
self.init_agent()
def __init__(self, size, p_hunter, p_prey,
prey_reward = 1, stuck_penalty = 1,
death_penalty = 1, p_resurrection = 0.2,
agent_max_age = 10, agent_range = 2, num_actions = 5,
same = True, lock = None, name=None,
max_iteration = 5000, boundary=False):
(H, W) = self.size = size
input_size = (2*agent_range + 1) ** 2
self.boundary_exists = boundary
if lock:
self.lock = lock
else:
self.lock = Lock()
self.A_mind = Mind(input_size, num_actions, self.lock, Queue())
self.B_mind = Mind(input_size, num_actions, self.lock, Queue())
self.max_iteration = max_iteration
self.lock = lock
if same:
weights = self.A_mind.network.state_dict()
self.B_mind.network.load_state_dict(weights)
self.p_prey = p_prey
self.p_resurrection = p_resurrection
assert 1 - 2*p_hunter - p_prey > 0, 'Free space probability is less than one'
self.probs = np.array([p_hunter, 1 - 2*p_hunter - p_prey, p_hunter, p_prey])
self.prey_reward = prey_reward
self.stuck_penalty = stuck_penalty
self.death_penalty = death_penalty
self.agent_max_age = agent_max_age
self.hzn = agent_range
self.vals = [-1, 0, 1, 2]
self.names_to_vals = {"void": -2, "A": -1, "free": 0, "B": 1, "prey": 2}
self.vals_to_names = {v: k for k, v in self.names_to_vals.items()}
self.vals_to_index = {-1: 0, 0: 1, 1: 2, 2: 3}
self.num_grids = size[0] * size[1]
self.id_to_type = {}
self.id_to_lives = {}
self.crystal = np.zeros((max_iteration, H, W, 3)) # type, age, id
self.history = []
self.id_track = []
self.records = []
self.args = [self.prey_reward,
self.stuck_penalty,
self.death_penalty,
self.agent_max_age,
self.hzn]
if name:
self.name = name
else:
self.name = abs(hash(tuple([self] + self.args)))
if not os.path.isdir(str(self.name)):
os.mkdir(str(self.name))
os.mkdir(str(self.name)+'/episodes')
self.map, self.agents, self.loc_to_agent, self.id_to_agent = self._generate_map()
self._set_initial_states()
self.mask = self._get_mask()
self.crystal = np.zeros((max_iteration, H, W, 4)) # type, age, tr, id
self.iteration = 0
else:
assert False, "There exists an experiment with this name."
class Environment:
def __init__(self, size, p_hunter, p_prey,
prey_reward = 1, stuck_penalty = 1,
death_penalty = 1, p_resurrection = 0.2,
agent_max_age = 10, agent_range = 2, num_actions = 5,
same = True, lock = None, name=None,
max_iteration = 5000, boundary=False):
(H, W) = self.size = size
input_size = (2*agent_range + 1) ** 2
self.boundary_exists = boundary
if lock:
self.lock = lock
else:
self.lock = Lock()
self.A_mind = Mind(input_size, num_actions, self.lock, Queue())
self.B_mind = Mind(input_size, num_actions, self.lock, Queue())
self.max_iteration = max_iteration
self.lock = lock
if same:
weights = self.A_mind.network.state_dict()
self.B_mind.network.load_state_dict(weights)
self.p_prey = p_prey
self.p_resurrection = p_resurrection
assert 1 - 2*p_hunter - p_prey > 0, 'Free space probability is less than one'
self.probs = np.array([p_hunter, 1 - 2*p_hunter - p_prey, p_hunter, p_prey])
self.prey_reward = prey_reward
self.stuck_penalty = stuck_penalty
self.death_penalty = death_penalty
self.agent_max_age = agent_max_age
self.hzn = agent_range
self.vals = [-1, 0, 1, 2]
self.names_to_vals = {"void": -2, "A": -1, "free": 0, "B": 1, "prey": 2}
self.vals_to_names = {v: k for k, v in self.names_to_vals.items()}
self.vals_to_index = {-1: 0, 0: 1, 1: 2, 2: 3}
self.num_grids = size[0] * size[1]
self.id_to_type = {}
self.id_to_lives = {}
self.crystal = np.zeros((max_iteration, H, W, 3)) # type, age, id
self.history = []
self.id_track = []
self.records = []
self.args = [self.prey_reward,
self.stuck_penalty,
self.death_penalty,
self.agent_max_age,
self.hzn]
if name:
self.name = name
else:
self.name = abs(hash(tuple([self] + self.args)))
if not os.path.isdir(str(self.name)):
os.mkdir(str(self.name))
os.mkdir(str(self.name)+'/episodes')
self.map, self.agents, self.loc_to_agent, self.id_to_agent = self._generate_map()
self._set_initial_states()
self.mask = self._get_mask()
self.crystal = np.zeros((max_iteration, H, W, 4)) # type, age, tr, id
self.iteration = 0
else:
assert False, "There exists an experiment with this name."
def configure(self, prey, penalty, age):
self.prey_reward = prey
self.stuck_penalty = penalty
self.agent_max_age = age
def get_agents(self):
return self.agents
def get_map(self):
return self.map.copy()
def move(self, agent):
(i, j) = loc = agent.get_loc()
(i_n, j_n) = to = agent.get_decision()
self.map[i, j] = 0
self.map[i_n, j_n] = agent.get_type()
agent.set_loc(to)
self.loc_to_agent[to] = agent
del self.loc_to_agent[loc]
def step(self, agent, by):
if agent.is_alive() and agent.get_time_remaining() == 0:
rew = self.kill(agent)
elif agent.is_alive():
(i, j) = agent.get_loc() # current location
assert self.loc_to_agent[(i, j)]
(di, dj) = by
(i_n, j_n) = self._add((i, j), (di, dj)) #new location
agent.set_decision((i_n, j_n))
if self.map[i_n, j_n] == self.names_to_vals["free"]:
rew = self.on_free(agent)
assert rew != None
elif self.map[i_n, j_n] == self.names_to_vals["prey"]:
prey = self.loc_to_agent[(i_n, j_n)]
if agent.get_type() in [-1, 1]:
rew = self.on_prey(agent, prey)
else:
rew = self.on_same(agent, prey)
assert rew != None
elif self.map[i_n, j_n] * agent.get_type() == -1:
opponent = self.loc_to_agent[(i_n, j_n)]
rew = self.on_opponent(agent, opponent)
assert rew != None
elif di == 0 and dj == 0:
rew = self.on_still(agent)
else:
other_agent = self.loc_to_agent[(i_n, j_n)]
rew = self.on_same(agent, other_agent)
assert rew != None
done = False
self.update_agent(agent, rew, done)
agent.clear_decision()
else:
resurrected = self.resurrect(agent)
rew = 0
if resurrected and agent.get_type() in [-1, 1]:
assert agent.get_time_remaining() == self.agent_max_age, agent.get_time_remaining()
assert agent.get_loc() != None, "wtf?"
return rew
def on_free(self, agent):
raise NotImplementedError('Please implement what to'\
'do when agent steps on a free grid.')
def on_still(self, agent):
raise NotImplementedError('Please implement what to'\
'do when agent stands still.')
def on_prey(self, agent, prey):
raise NotImplementedError('Please implement what to'\
'do when agent encounters a prey.')
def on_obstacle(self, agent):
raise NotImplementedError('Please implement what to'\
'do when agent encounters an obstacle.')
def on_same(self, agent, other):
raise NotImplementedError('Please implement what to'\
'do when agent encounters its same kind.')
def on_opponent(self, agent, opponent):
raise NotImplementedError('Please implement what to'\
'do when agent encounters a different kind.')
def kill(self, victim, killer=None):
raise NotImplementedError('Please implement what to'\
'do when victim dies (with or wo killer).')
def resurrect(self, agent):
resurrected = False
if np.random.random() < self.p_resurrection:
locs = [(a, b) for a in range(len(self.map))
for b in range(len(self.map[0]))
if self.map[a,b] == 0]
idx = np.random.choice(range(len(locs)))
(i, j) = loc = locs[idx]
self.map[i, j] = agent.get_type()
self.loc_to_agent[loc] = agent
agent.respawn(loc)
agent.set_current_state(self.get_agent_state(agent))
resurrected = True
return resurrected
def update_agent(self, agent, rew, done):
state = self.get_agent_state(agent)
agent.set_next_state(state)
name = self.vals_to_names[agent.get_type()]
agent.update(rew, done)
return rew
def update(self):
self.iteration += 1
self.history.append(self.map.copy())
self.A_mind.train(self.names_to_vals["A"])
self.B_mind.train(self.names_to_vals["B"])
a_ages = []
a_ids = []
b_ages = []
b_ids = []
id_track = np.zeros(self.map.shape)
self.deads = []
for agent in self.agents:
typ = agent.get_type()
age = agent.get_age()
idx = agent.get_id()
if agent.is_alive():
i, j = agent.get_loc()
tr = agent.get_time_remaining()
id_track[i, j] = idx
self.crystal[self.iteration - 1, i, j] = [typ, age, tr, idx]
else:
self.deads.append([agent.get_type(), agent.get_id()])
type = agent.get_type()
if type == self.names_to_vals["A"]:
a_ages.append(str(age))
a_ids.append(str(idx))
else:
b_ages.append(str(age))
b_ids.append(str(idx))
self.id_track.append(id_track)
a_ages = " ".join(a_ages)
b_ages = " ".join(b_ages)
a_ids = " ".join(a_ids)
b_ids = " ".join(b_ids)
with open("%s/episodes/a_age.csv" % self.name, "a") as f:
f.write("%s, %s, %s\n" % (self.iteration, a_ages, a_ids))
with open("%s/episodes/b_age.csv" % self.name, "a") as f:
f.write("%s, %s, %s\n" % (self.iteration, b_ages, b_ids))
if self.iteration == self.max_iteration - 1:
A_losses = self.A_mind.get_losses()
B_losses = self.B_mind.get_losses()
np.save("%s/episodes/a_loss.npy" % self.name, np.array(A_losses))
np.save("%s/episodes/b_loss.npy" % self.name, np.array(B_losses))
def shuffle(self):
map = np.zeros(self.size)
loc_to_agent = {}
locs = [(i, j) for i in range(self.map.shape[0]) for j in range(self.map.shape[1]) if self.map[i, j] == 0]
random.shuffle(locs)
id_track = np.zeros(self.map.shape)
for i, agent in enumerate(self.agents):
loc = locs[i]
agent.respawn(loc)
loc_to_agent[loc] = agent
map[loc] = agent.get_type()
id_track[loc] = agent.get_id()
self.map, self.loc_to_agent = map, loc_to_agent
self._set_initial_states()
self.history = [map.copy()]
self.id_track = [id_track]
self.records = []
self.iteration = 0
def record(self, rews):
self.records.append(rews)
def save(self, episode):
f = gzip.GzipFile('%s/crystal.npy.gz' % self.name, "w")
np.save(f, self.crystal)
f.close()
def save_agents(self):
self.lock.acquire()
pickle.dump(self.agents, open("agents/agent_%s.p" % (self.name), "wb" ))
self.lock.release()
def get_agent_state(self, agent):
hzn = self.hzn
i, j = agent.get_loc()
fov = np.zeros((2 * hzn + 1, 2 * hzn + 1)) - 2
if self.boundary_exists:
start_i, end_i, start_j, end_j = 0, 2 * hzn + 1, 0, 2 * hzn + 1
if i < hzn:
start_i = hzn - i
elif i + hzn - self.size[0] + 1 > 0:
end_i = (2 * hzn + 1) - (i + hzn - self.size[0] + 1)
if j < hzn:
start_j = hzn - j
elif j + hzn - self.size[1] + 1 > 0:
end_j = (2 * hzn + 1) - (j + hzn - self.size[1] + 1)
i_upper = min(i + hzn + 1, self.size[0])
i_lower = max(i - hzn, 0)
j_upper = min(j + hzn + 1, self.size[1])
j_lower = max(j - hzn, 0)
fov[start_i: end_i, start_j: end_j] = self.map[i_lower: i_upper, j_lower: j_upper].copy()
else:
for di in range(-hzn, hzn+1):
for dj in range(-hzn, hzn+1):
fov[hzn + di, hzn + dj] = self.map[(i+di) % self.size[0], (j+dj) % self.size[1]]
fov[hzn, hzn] = agent.get_type()
return fov
def _to_csv(self, episode):
with open("episodes/%s_%s.csv" % (episode, self.name), 'w') as f:
f.write(', '.join(self.records[0].keys()) + '\n')
proto = ", ".join(['%.3f' for _ in range(len(self.records[0]))]) + '\n'
for rec in self.records:
f.write(proto % tuple(rec.values()))
def _add(self, fr, by):
i, j = fr
di, dj = by
if self.boundary_exists:
i_n = min(max(i + di, 0), self.size[0] - 1)
j_n = min(max(j + dj, 0), self.size[1] - 1)
else:
i_n = (i + di) % self.size[0]
j_n = (j + dj) % self.size[1]
return (i_n, j_n)
def _get_mask(self):
mask = []
for i, row in enumerate(self.map):
foo = []
for j, col in enumerate(row):
foo.append((-1) ** (i + j))
mask.append(foo)
return np.array(mask)
def _generate_map(self):
map = np.zeros(self.size)
loc_to_agent = {}
id_to_agent = {}
agents = []
idx = 0
for i, row in enumerate(map):
for j, col in enumerate(row):
val = np.random.choice(self.vals, p=self.probs)
if not val == self.names_to_vals["free"]:
if val == self.names_to_vals["A"]:
mind = self.A_mind
elif val == self.names_to_vals["B"]:
mind = self.B_mind
else:
assert False, 'Error'
agent = Agent(idx, (i, j), val, mind, self.probs[1], self.agent_max_age)
loc_to_agent[(i, j)] = agent
id_to_agent[idx] = agent
agents.append(agent)
idx += 1
map[i, j] = val
return map, agents, loc_to_agent, id_to_agent
def predefined_initialization(self, file):
with open(file) as f:
for i, line in enumerate(f):
if not i:
keys = [key.strip() for key in line.rstrip().split(',')]
line.rstrip().split(',')
def _set_initial_states(self):
for agent in self.agents:
state = self.get_agent_state(agent)
agent.set_current_state(state)
def _count(self, arr):
cnt = np.zeros(len(self.vals))
arr = arr.reshape(-1)
for elem in arr:
if elem in self.vals_to_index:
cnt[self.vals_to_index[elem]] += 1
return cnt
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.feature_extraction.text import TfidfTransformer
from nltk.corpus import stopwords
import numpy as np
import numpy.linalg as LA
from mind import Mind
m = Mind()
m.analyzeShortTermMemory()
#train_set = ["The sky is blue.", "The sun is bright."] #Documents
train_set = list(m.botsMemory.keys())
test_set = ["hello how are you"] #Query
vectorizer = CountVectorizer(stop_words = None)
#print vectorizer
transformer = TfidfTransformer()
#print transformer
trainVectorizerArray = vectorizer.fit_transform(train_set).toarray()
testVectorizerArray = vectorizer.transform(test_set).toarray()
print 'Fit Vectorizer to train set', trainVectorizerArray
print 'Transform Vectorizer to test set', testVectorizerArray
cx = lambda a, b : round(np.inner(a, b)/(LA.norm(a)*LA.norm(b)), 6)
for vector in trainVectorizerArray:
#print vector
for testV in testVectorizerArray:
#print testV
cosine = cx(vector, testV)
import torch
data = get_data('../data/samples.txt')
mid, seq, avid, label, lens = collect_data(batch_data=data)
dataset = MindDataset(mid, seq, avid, label, lens)
dataloader = DataLoader(dataset, batch_size=32)
dim = 8
max_len = 20
input_units = 20
output_units = 20
iteration = 1
max_K = 3
p = 2
model = Mind(dim, max_len, input_units, output_units, iteration, max_K, p)
optimizer = torch.optim.SGD(model.parameters(), lr=0.01, momentum=0.9)
model.train()
train_iteration = 3
for i in range(train_iteration):
losses = []
for data in dataloader:
mid, seq, label_avid, label, lens = data
output = model(mid, seq, label_avid, lens)
output = output.squeeze(dim=-1).float()
label = label.float()
loss = F.binary_cross_entropy_with_logits(output, label)
losses.append(loss.data.item())
def main():
from mind import Mind
minecraft_mind = Mind()
minecraft_mind.learning_loop()