def load_rtneat(self, location , pop, team=constants.OBJECT_TYPE_TEAM_0):
location = os.path.relpath("/") + location
if os.path.exists(location):
OpenNero.set_ai("rtneat-%s" % team, OpenNero.RTNEAT(
str(location), "data/ai/neat-params.dat",
constants.pop_size,
OpenNero.get_environment().agent_info.reward))
def deploy(self, ai='rtneat', team=constants.OBJECT_TYPE_TEAM_0):
OpenNero.disable_ai()
if ai == 'rtneat':
OpenNero.set_ai('rtneat-%s' % team, None)
self.environment.remove_all_agents(team)
for _ in range(constants.pop_size):
self.spawnAgent(ai=ai, team=team)
OpenNero.enable_ai()
def deploy(self, ai='rtneat', team=constants.OBJECT_TYPE_TEAM_0):
OpenNero.disable_ai()
if ai == 'rtneat':
OpenNero.set_ai('rtneat-%s' % team, None)
self.environment.remove_all_agents(team)
for _ in range(constants.pop_size):
self.spawnAgent(ai=ai, team=team)
OpenNero.enable_ai()
def load_rtneat(self, location, pop, team=constants.OBJECT_TYPE_TEAM_0):
location = os.path.relpath("/") + location
if os.path.exists(location):
OpenNero.set_ai(
"rtneat-%s" % team,
OpenNero.RTNEAT(str(location), "data/ai/neat-params.dat",
constants.pop_size,
OpenNero.get_environment().agent_info.reward))
def __init__(self):
"""
Create the environment
"""
OpenNero.Environment.__init__(self)
self.curr_id = 0
self.max_steps = 20
self.MAX_DIST = math.hypot(constants.XDIM, constants.YDIM)
self.states = {}
self.teams = {}
self.script = 'Hw5/menu.py'
abound = OpenNero.FeatureVectorInfo() # actions
sbound = OpenNero.FeatureVectorInfo() # sensors
rbound = OpenNero.FeatureVectorInfo() # rewards
# actions
abound.add_continuous(
-1, 1
) # forward/backward speed (gets multiplied by constants.MAX_MOVEMENT_SPEED)
abound.add_continuous(
-constants.MAX_TURN_RADIANS,
constants.MAX_TURN_RADIANS) # left/right turn (in radians)
# sensor dimensions
for a in range(constants.N_SENSORS):
sbound.add_continuous(0, 1)
# Rewards
# the enviroment returns the raw multiple dimensions of the fitness as
# they get each step. This then gets combined into, e.g. Z-score, by
# the ScoreHelper in order to calculate the final rtNEAT-fitness
for f in constants.FITNESS_DIMENSIONS:
# we don't care about the bounds of the individual dimensions
rbound.add_continuous(-sys.float_info.max,
sys.float_info.max) # range for reward
# initialize the rtNEAT algorithm parameters
# input layer has enough nodes for all the observations plus a bias
# output layer has enough values for all the actions
# population size matches ours
# 1.0 is the weight initialization noise
rtneat = OpenNero.RTNEAT("data/ai/neat-params.dat",
OpenNero.Population(), constants.pop_size, 1)
key = "rtneat-%s" % constants.OBJECT_TYPE_TEAM_0
OpenNero.set_ai(key, rtneat)
print "get_ai(%s): %s" % (key, OpenNero.get_ai(key))
# set the initial lifetime
lifetime = module.getMod().lt
rtneat.set_lifetime(lifetime)
print 'rtNEAT lifetime:', lifetime
self.agent_info = OpenNero.AgentInitInfo(sbound, abound, rbound)
def __init__(self):
"""
Create the environment
"""
OpenNero.Environment.__init__(self)
self.curr_id = 0
self.max_steps = 20
self.MAX_DIST = math.hypot(constants.XDIM, constants.YDIM)
self.states = {}
self.teams = {}
self.script = 'Hw5/menu.py'
abound = OpenNero.FeatureVectorInfo() # actions
sbound = OpenNero.FeatureVectorInfo() # sensors
rbound = OpenNero.FeatureVectorInfo() # rewards
# actions
abound.add_continuous(-1, 1) # forward/backward speed (gets multiplied by constants.MAX_MOVEMENT_SPEED)
abound.add_continuous(-constants.MAX_TURN_RADIANS, constants.MAX_TURN_RADIANS) # left/right turn (in radians)
# sensor dimensions
for a in range(constants.N_SENSORS):
sbound.add_continuous(0, 1);
# Rewards
# the enviroment returns the raw multiple dimensions of the fitness as
# they get each step. This then gets combined into, e.g. Z-score, by
# the ScoreHelper in order to calculate the final rtNEAT-fitness
for f in constants.FITNESS_DIMENSIONS:
# we don't care about the bounds of the individual dimensions
rbound.add_continuous(-sys.float_info.max, sys.float_info.max) # range for reward
# initialize the rtNEAT algorithm parameters
# input layer has enough nodes for all the observations plus a bias
# output layer has enough values for all the actions
# population size matches ours
# 1.0 is the weight initialization noise
rtneat = OpenNero.RTNEAT("data/ai/neat-params.dat",
constants.N_SENSORS,
constants.N_ACTIONS,
constants.pop_size,
1.0,
rbound, False)
key = "rtneat-%s" % constants.OBJECT_TYPE_TEAM_0
OpenNero.set_ai(key, rtneat)
print "get_ai(%s): %s" % (key, OpenNero.get_ai(key))
# set the initial lifetime
lifetime = module.getMod().lt
rtneat.set_lifetime(lifetime)
print 'rtNEAT lifetime:', lifetime
self.agent_info = OpenNero.AgentInitInfo(sbound, abound, rbound)
def start_rtneat(self, pop_size):
" start the rtneat learning demo "
OpenNero.disable_ai()
#self.environment = RoombaEnvironment(constants.XDIM, constants.YDIM, self)
#set_environment(self.environment)
#self.reset_sandbox()
# Create RTNEAT object
rbound = OpenNero.FeatureVectorInfo()
rbound.add_continuous(-sys.float_info.max, sys.float_info.max)
rtneat = OpenNero.RTNEAT("data/ai/neat-params.dat", 2, 1, pop_size, 1.0, rbound, False)
rtneat.set_weight(0,1)
OpenNero.set_ai("rtneat",rtneat)
OpenNero.enable_ai()
self.distribute_bots(pop_size, "data/shapes/roomba/RoombaRTNEAT.xml")
def start_rtneat(self, team=constants.OBJECT_TYPE_TEAM_0):
# initialize the rtNEAT algorithm parameters
# input layer has enough nodes for all the observations plus a bias
# output layer has enough values for all the actions
# population size matches ours
# 1.0 is the weight initialization noise
rtneat = OpenNero.RTNEAT("data/ai/neat-params.dat",
constants.N_SENSORS + 1,
constants.N_ACTIONS, constants.pop_size, 1.0,
rtneat_rewards(), False)
key = "rtneat-%s" % team
OpenNero.set_ai(key, rtneat)
print "get_ai(%s): %s" % (key, OpenNero.get_ai(key))
rtneat.set_lifetime(self.environment.lifetime)
def start_rtneatq(self, team=constants.OBJECT_TYPE_TEAM_0):
# initialize the rtNEAT+Q algorithm parameters
# input layer has enough nodes for all the observations plus a bias
# output layer has enough values for all the wires
# population size matches ours
# 1.0 is the weight initialization noise
rtneatq = OpenNero.RTNEAT("data/ai/neat-params.dat",
constants.N_SENSORS+1,
constants.N_ACTION_CANDIDATES * (constants.N_ACTIONS + 1),
constants.pop_size,
1.0,
rtneat_rewards(),
False)
key = "rtneatq-%s" % team
OpenNero.set_ai(key, rtneatq)
print "get_ai(%s): %s" % (key, OpenNero.get_ai(key))
rtneatq.set_lifetime(self.environment.lifetime)
class NeroModule:
def __init__(self):
self.environment = None
self.agent_id = None
self.flag_loc = None
self.flag_id = None
self.set_speedup(constants.DEFAULT_SPEEDUP)
x = constants.XDIM / 2.0
y = constants.YDIM / 3.0
self.spawn_x = {}
self.spawn_y = {}
self.set_spawn(x, y, constants.OBJECT_TYPE_TEAM_0)
self.set_spawn(x, 2 * y, constants.OBJECT_TYPE_TEAM_1)
# Bounds for sensors in neural network and advice language. These bounds are
# used to convert sensor values between network and advice.
self.sbounds_network = OpenNero.FeatureVectorInfo()
self.sbounds_advice = OpenNero.FeatureVectorInfo()
# Networks have better learning bias when cube sensors produce values in the
# range [-1, 1], but the range [0, 1] is more intuitive in the advice
# language. Wall sensors use the same range [0, 1] for both network and advice.
# The sense() method in the ForageEnvironment class use these network bounds
# to scale the sensor values.
for i in range(constants.N_SENSORS):
self.sbounds_network.add_continuous(0, 1)
self.sbounds_advice.add_continuous(0, 1)
# The last sensor is the bias, which always takes the value 1 (upper bound).
self.sbounds_network.add_continuous(0, 1)
self.sbounds_advice.add_continuous(0, 1)
print 'sbounds_network', self.sbounds_network
def setup_map(self):
"""
setup the test environment
"""
OpenNero.disable_ai()
if self.environment:
error("Environment already created")
return
# create the environment - this also creates the rtNEAT object
self.environment = self.create_environment()
OpenNero.set_environment(self.environment)
# world walls
height = constants.HEIGHT + constants.OFFSET
common.addObject("data/shapes/cube/Cube.xml",
OpenNero.Vector3f(constants.XDIM / 2, 0, height),
OpenNero.Vector3f(0, 0, 90),
scale=OpenNero.Vector3f(constants.WIDTH,
constants.XDIM,
constants.HEIGHT * 2),
label="World Wall0",
type=constants.OBJECT_TYPE_OBSTACLE)
common.addObject("data/shapes/cube/Cube.xml",
OpenNero.Vector3f(0, constants.YDIM / 2, height),
OpenNero.Vector3f(0, 0, 0),
scale=OpenNero.Vector3f(constants.WIDTH,
constants.YDIM,
constants.HEIGHT * 2),
label="World Wall1",
type=constants.OBJECT_TYPE_OBSTACLE)
common.addObject("data/shapes/cube/Cube.xml",
OpenNero.Vector3f(constants.XDIM, constants.YDIM / 2,
height),
OpenNero.Vector3f(0, 0, 0),
scale=OpenNero.Vector3f(constants.WIDTH,
constants.YDIM,
constants.HEIGHT * 2),
label="World Wall2",
type=constants.OBJECT_TYPE_OBSTACLE)
common.addObject("data/shapes/cube/Cube.xml",
OpenNero.Vector3f(constants.XDIM / 2, constants.YDIM,
height),
OpenNero.Vector3f(0, 0, 90),
scale=OpenNero.Vector3f(constants.WIDTH,
constants.XDIM,
constants.HEIGHT * 2),
label="World Wall3",
type=constants.OBJECT_TYPE_OBSTACLE)
# Add an obstacle wall in the middle
common.addObject("data/shapes/cube/Cube.xml",
OpenNero.Vector3f(constants.XDIM / 2,
constants.YDIM / 2, height),
OpenNero.Vector3f(0, 0, 90),
scale=OpenNero.Vector3f(constants.WIDTH,
constants.YDIM / 4,
constants.HEIGHT * 2),
label="World Wall4",
type=constants.OBJECT_TYPE_OBSTACLE)
# Add some trees
for i in (0.25, 0.75):
for j in (0.25, 0.75):
# don't collide with trees - they are over 500 triangles each
common.addObject("data/shapes/tree/Tree.xml",
OpenNero.Vector3f(i * constants.XDIM,
j * constants.YDIM,
constants.OFFSET),
OpenNero.Vector3f(0, 0, 0),
scale=OpenNero.Vector3f(1, 1, 1),
label="Tree %d %d" % (10 * i, 10 * j),
type=constants.OBJECT_TYPE_LEVEL_GEOM)
# collide with their trunks represented with cubes instead
common.addObject(
"data/shapes/cube/Cube.xml",
OpenNero.Vector3f(i * constants.XDIM, j * constants.YDIM,
constants.OFFSET),
OpenNero.Vector3f(0, 0, 0),
scale=OpenNero.Vector3f(1, 1, constants.HEIGHT),
type=constants.OBJECT_TYPE_OBSTACLE)
# Add the surrounding Environment
common.addObject("data/terrain/NeroWorld.xml",
OpenNero.Vector3f(constants.XDIM / 2,
constants.YDIM / 2, 0),
scale=OpenNero.Vector3f(1.2, 1.2, 1.2),
label="NeroWorld",
type=constants.OBJECT_TYPE_LEVEL_GEOM)
return True
def create_environment(self):
return NeroEnvironment.NeroEnvironment()
def remove_flag(self):
if self.flag_id:
common.removeObject(self.flag_id)
def change_flag(self, loc):
if self.flag_id:
common.removeObject(self.flag_id)
self.flag_loc = OpenNero.Vector3f(*loc)
self.flag_id = common.addObject("data/shapes/cube/BlueCube.xml",
self.flag_loc,
label="Flag",
scale=OpenNero.Vector3f(1, 1, 10),
type=constants.OBJECT_TYPE_FLAG)
def place_basic_turret(self, loc):
return common.addObject("data/shapes/character/steve_basic_turret.xml",
OpenNero.Vector3f(*loc),
type=constants.OBJECT_TYPE_TEAM_1)
#The following is run when one of the Deploy buttons is pressed
def deploy(self, ai='rtneat', team=constants.OBJECT_TYPE_TEAM_0):
OpenNero.disable_ai()
if ai == 'rtneat':
OpenNero.set_ai('rtneat-%s' % team, None)
self.environment.remove_all_agents(team)
for _ in range(constants.pop_size):
self.spawnAgent(ai=ai, team=team)
OpenNero.enable_ai()
#The following is run when the Save button is pressed
def save_team(self, location, team=constants.OBJECT_TYPE_TEAM_0):
# if there are rtneat agents in the environment, save them as a group.
rtneat = OpenNero.get_ai("rtneat-%s" % team)
if rtneat:
location = rtneat.save_population(str(location))
# then, check whether there are any qlearning agents, and save them.
with open(location, 'a') as handle:
for agent in self.environment.teams[team]:
if agent.group == 'Agent' and agent.ai == 'qlearning':
handle.write('\n\n%s' % agent.to_string())
if hasattr(agent, 'stats'):
handle.write('\n\n%s' % agent.stats())
#The following is run when the Load button is pressed
def load_team(self, location, team=constants.OBJECT_TYPE_TEAM_0):
OpenNero.disable_ai()
self.environment.remove_all_agents(team)
if not os.path.exists(location):
print location, 'does not exist, cannot load population'
return
# parse out different agents from the population file.
contents = ''
try:
try:
handle = gzip.open(location)
contents = handle.read()
finally:
handle.close()
except Exception, e:
with open(location) as handle:
contents = handle.read()
if not contents:
print 'cannot read', location, 'skipping'
return
rtneat, qlearning, stats = self._split_population(
contents.splitlines(True))
print 'qlearning agents:', qlearning.count('Approximator')
# load any qlearning agents first, subtracting them from the population
# size that rtneat will need to manage. since we cannot deserialize an
# agent's state until after it's been added to the world, we put the
# serialized chunk for the agent into a map, then NeroEnvironment#step
# takes care of the deserialization.
pop_size = constants.pop_size
if qlearning.strip():
for chunk in re.split(r'\n\n+', qlearning):
if not chunk.strip():
continue
id = self.spawnAgent(ai='qlearning', team=team)
self.environment.agents_to_load[id] = chunk
pop_size -= 1
if pop_size == 0:
break
print 'rtneat agents:', rtneat.count('genomeend')
# load any rtneat agents from the file, as a group.
if pop_size > 0 and rtneat.strip():
tf = tempfile.NamedTemporaryFile(delete=False)
tf.write(rtneat)
tf.close()
print tf.name
OpenNero.set_ai(
"rtneat-%s" % team,
OpenNero.RTNEAT(tf.name, "data/ai/neat-params.dat", pop_size,
rtneat_rewards(), False))
os.unlink(tf.name)
while pop_size > 0:
self.spawnAgent(ai='rtneat', team=team)
pop_size -= 1
OpenNero.enable_ai()
def stop_training(self):
OpenNero.set_ai('rtneat-%s' % self.team_type, None)
def start_training(self):
OpenNero.set_ai('rtneat-%s' % self.team_type, self.rtneat)