def grid_world1_trp(exp_id=4, path="./Results/gridworld1"):
opt = {}
opt["exp_id"] = exp_id
opt["path"] = path
opt["checks_per_policy"] = 10
opt["max_steps"] = 150000
opt["num_policy_checks"] = 20
noise = 0.1
exp = 0.3
discretization = 20
# Domain:
maze = os.path.join(ConsumableGridWorld.default_map_dir, '10x7-ACC2011.txt')
domain = ConsumableGridWorldIRL([(7,5), (1,2)],
mapname=maze,
encodingFunction= lambda x: ConsumableGridWorldIRL.stateVisitEncoding(x,[(7,5)]),
binary=True,
noise=noise)
#domain = Pinball(noise=0.3)
# Representation
representation = Tabular(domain, discretization=discretization)
# Policy
policy = eGreedy(representation, epsilon=0.3)
d = GoalPathPlanner(domain, representation,policy)
trajs = d.generateTrajectories(N=5)
a = TransitionStateClustering(window_size=2)
for t in trajs:
N = len(t)
demo = np.zeros((N,2))
for i in range(0,N):
demo[i,:] = t[i][0:2]
a.addDemonstration(demo)
a.fit(normalize=False, pruning=0.5)
ac = [(round(a.means_[0][0]),round(a.means_[0][1])) for a in a.model]
print ac
#reinitialize
domain = ConsumableGridWorldIRL([(7,5), (1,2)],
mapname=maze,
encodingFunction= lambda x: ConsumableGridWorldIRL.statePassageEncoding(x,ac,5), noise=noise)
representation = IncrementalTabular(domain, discretization=discretization)
policy = eGreedy(representation, epsilon=0.3)
opt["agent"] = Q_Learning(representation=representation, policy=policy,
discount_factor=domain.discount_factor,
initial_learn_rate=0.1,
learn_rate_decay_mode="boyan", boyan_N0=100,
lambda_=0.)
opt["domain"] = domain
experiment = Experiment(**opt)
experiment.run(visualize_steps=False,
visualize_learning=False,
visualize_performance=0)
experiment.save()
return np.max(experiment.result["return"]),np.sum(experiment.result["return"])
def createSlidingDomain(self, k):
return ConsumableGridWorldIRL(
self.env_template["consumable"],
mapname=self.env_template["map"],
encodingFunction=lambda x: ConsumableGridWorldIRL.
slidingWindowEncoding(x, k),
noise=self.env_template["noise"])
def runRewardIRL(self,N=5):
opt = deepcopy(self.opt_template)
dist = self.getIRLTDist(self.env_template["consumable"],N=N)
bdist = self.getIRLDist(N=N, rand=True)
dist = [d-bdist for d in dist]
print dist
domain = self.createStateDomain(waypoints=self.env_template["consumable"],
rewardFunction=lambda x,y,z,w: ConsumableGridWorldIRL.rewardIRL(x,y,z,w,dist,self.env_template["consumable"]))
opt["domain"] = domain
representation = IncrementalTabular(domain, discretization=self.env_template["discretization"])
policy = eGreedy(representation, epsilon=self.env_template["exp"])
opt["agent"] = Q_Learning(representation=representation, policy=policy,
discount_factor=domain.discount_factor,
initial_learn_rate=0.1,
learn_rate_decay_mode="boyan", boyan_N0=100,
lambda_=0.)
experiment = Experiment(**opt)
experiment.run(visualize_steps=False,
performance_domain = self.createStateDomain(self.env_template["consumable"]),
visualize_learning=False,
visualize_performance=0)
experiment.save()
return np.max(experiment.result["return"]),np.sum(experiment.result["return"])
def runTIRL(self, N=5, w=2, pruning=0.5):
opt = deepcopy(self.opt_template)
dist = self.getIRLDist(N=N)
ac = self.getTSCWaypoints(N, w, pruning)
domain = self.createStateDomain(
waypoints=ac,
rewardFunction=lambda x, y, z, w: ConsumableGridWorldIRL.rewardIRL(
x, y, z, w, dist))
opt["domain"] = domain
representation = IncrementalTabular(
domain, discretization=self.env_template["discretization"])
policy = eGreedy(representation, epsilon=self.env_template["exp"])
opt["agent"] = Q_Learning(representation=representation,
policy=policy,
discount_factor=domain.discount_factor,
initial_learn_rate=0.1,
learn_rate_decay_mode="boyan",
boyan_N0=100,
lambda_=0.)
experiment = Experiment(**opt)
experiment.run(visualize_steps=False,
performance_domain=self.createStateDomain(
waypoints=self.env_template["consumable"]),
visualize_learning=False,
visualize_performance=0)
experiment.save()
return np.max(experiment.result["return"]), np.sum(
experiment.result["return"])
def createStateDomain(self,waypoints, rewardFunction=None):
return ConsumableGridWorldIRL(self.env_template["consumable"],
mapname=self.env_template["map"],
encodingFunction= lambda x: ConsumableGridWorldIRL.stateVisitEncoding(x,waypoints),
rewardFunction = rewardFunction,
noise=self.env_template["noise"],
binary=True)
def createMarkovDomain(self, k=1, rewardFunction=None):
return ConsumableGridWorldIRL(self.env_template["consumable"],
mapname=self.env_template["map"],
encodingFunction= lambda x: ConsumableGridWorldIRL.allMarkovEncoding(x,k),
rewardFunction = rewardFunction,
noise=self.env_template["noise"],
binary=True)
def createPassageDomain(self, selfwaypoints, k=5):
return ConsumableGridWorldIRL(
self.env_template["consumable"],
mapname=self.env_template["map"],
encodingFunction=lambda x: ConsumableGridWorldIRL.
statePassageEncoding(x, waypoints, k),
noise=self.env_template["noise"],
binary=True)
def grid_world1_reward(exp_id=2, path="./Results/gridworld1"):
opt = {}
opt["exp_id"] = exp_id
opt["path"] = path
opt["checks_per_policy"] = 10
opt["max_steps"] = 150000
opt["num_policy_checks"] = 20
noise = 0.1
exp = 0.3
discretization = 400
maze = os.path.join(ConsumableGridWorld.default_map_dir,
'10x7-ACC2011.txt')
domain = ConsumableGridWorldIRL(
[(7, 5), (1, 2)],
mapname=maze,
encodingFunction=lambda x: ConsumableGridWorldIRL.stateVisitEncoding(
x, [(7, 5)]),
noise=noise,
binary=True)
opt["domain"] = domain
# Representation
representation = Tabular(domain, discretization=discretization)
# Policy
policy = eGreedy(representation, epsilon=exp)
# Agent
opt["agent"] = Q_Learning(representation=representation,
policy=policy,
discount_factor=domain.discount_factor,
initial_learn_rate=0.1,
learn_rate_decay_mode="boyan",
boyan_N0=100,
lambda_=0.)
experiment = Experiment(**opt)
experiment.run(visualize_steps=False,
visualize_learning=False,
visualize_performance=0)
experiment.save()
return np.max(experiment.result["return"]), np.sum(
experiment.result["return"])
def createMarkovDomain(self, k=1, rewardFunction=None):
return ConsumableGridWorldIRL(
self.env_template["consumable"],
mapname=self.env_template["map"],
encodingFunction=lambda x: ConsumableGridWorldIRL.
allMarkovEncoding(x, k),
rewardFunction=rewardFunction,
noise=self.env_template["noise"],
binary=True)
def createStateDomain(self, waypoints, rewardFunction=None):
return ConsumableGridWorldIRL(
self.env_template["consumable"],
mapname=self.env_template["map"],
encodingFunction=lambda x: ConsumableGridWorldIRL.
stateVisitEncoding(x, waypoints),
rewardFunction=rewardFunction,
noise=self.env_template["noise"],
binary=True)
def grid_world1_reward(exp_id=2, path="./Results/gridworld1"):
opt = {}
opt["exp_id"] = exp_id
opt["path"] = path
opt["checks_per_policy"] = 10
opt["max_steps"] = 150000
opt["num_policy_checks"] = 20
noise = 0.1
exp = 0.3
discretization = 400
maze = os.path.join(ConsumableGridWorld.default_map_dir, '10x7-ACC2011.txt')
domain = ConsumableGridWorldIRL([(7,5), (1,2)],
mapname=maze,
encodingFunction= lambda x: ConsumableGridWorldIRL.stateVisitEncoding(x,[(7,5)]),
noise=noise,
binary=True)
opt["domain"] = domain
# Representation
representation = Tabular(domain, discretization=discretization)
# Policy
policy = eGreedy(representation, epsilon=exp)
# Agent
opt["agent"] = Q_Learning(representation=representation, policy=policy,
discount_factor=domain.discount_factor,
initial_learn_rate=0.1,
learn_rate_decay_mode="boyan", boyan_N0=100,
lambda_=0.)
experiment = Experiment(**opt)
experiment.run(visualize_steps=False,
visualize_learning=False,
visualize_performance=0)
experiment.save()
return np.max(experiment.result["return"]),np.sum(experiment.result["return"])
def gridworld1_tirl(exp_id=7, path="./Results/gridworld1"):
opt = {}
opt["exp_id"] = exp_id
opt["path"] = path
opt["checks_per_policy"] = 10
opt["max_steps"] = 150000
opt["num_policy_checks"] = 20
noise = 0.1
exp = 0.3
discretization = 400
# Domain:
maze = os.path.join(ConsumableGridWorld.default_map_dir,
'10x7-ACC2011.txt')
domain = ConsumableGridWorldIRL(
[(7, 5), (1, 2)],
mapname=maze,
encodingFunction=lambda x: ConsumableGridWorldIRL.stateVisitEncoding(
x, [(7, 5)]),
noise=noise,
binary=True)
#domain = Pinball(noise=0.3)
# Representation
representation = Tabular(domain, discretization=discretization)
# Policy
policy = eGreedy(representation, epsilon=0.3)
# Agent
opt["agent"] = Q_Learning(representation=representation,
policy=policy,
discount_factor=domain.discount_factor,
initial_learn_rate=0.1,
learn_rate_decay_mode="boyan",
boyan_N0=100,
lambda_=0.)
opt["checks_per_policy"] = 10
opt["max_steps"] = 150000
opt["num_policy_checks"] = 20
d = GoalPathPlanner(domain, representation, policy)
trajs = d.generateTrajectories(N=5)
dist = calculateStateDist((10, 7), trajs)
a = TransitionStateClustering(window_size=2)
for t in trajs:
N = len(t)
demo = np.zeros((N, 2))
for i in range(0, N):
demo[i, :] = t[i][0:2]
a.addDemonstration(demo)
a.fit(normalize=False, pruning=0.5)
dist = calculateStateDist((10, 7), trajs)
ac = discrete2DClustersToPoints(a.model, dist, radius=1)
# Policy reset
policy = eGreedy(representation, epsilon=0.3)
representation = Tabular(domain, discretization=discretization)
opt["agent"] = Q_Learning(representation=representation,
policy=policy,
discount_factor=domain.discount_factor,
initial_learn_rate=0.1,
learn_rate_decay_mode="boyan",
boyan_N0=100,
lambda_=0.)
domain = ConsumableGridWorldIRL(
[(7, 5), (1, 2)],
mapname=maze,
encodingFunction=lambda x: ConsumableGridWorldIRL.stateVisitEncoding(
x, [(7, 5)]),
rewardFunction=lambda x, y, z, w: ConsumableGridWorldIRL.tRewardIRL(
x, y, z, w, dist, [(7, 5)]),
noise=noise)
pdomain = ConsumableGridWorldIRL(
[(7, 5), (1, 2)],
mapname=maze,
encodingFunction=lambda x: ConsumableGridWorldIRL.stateVisitEncoding(
x, [(7, 5)]),
noise=noise)
opt["domain"] = domain
experiment = Experiment(**opt)
experiment.run(visualize_steps=False,
performance_domain=pdomain,
visualize_learning=False,
visualize_performance=0)
experiment.save()
return np.max(experiment.result["return"]), np.sum(
experiment.result["return"])
def gridworld1_rirl(exp_id=6, path="./Results/gridworld1"):
opt = {}
opt["exp_id"] = exp_id
opt["path"] = path
opt["checks_per_policy"] = 10
opt["max_steps"] = 150000
opt["num_policy_checks"] = 20
noise = 0.1
exp = 0.3
discretization = 400
# Domain:
maze = os.path.join(ConsumableGridWorld.default_map_dir,
'10x7-ACC2011.txt')
domain = ConsumableGridWorldIRL(
[(7, 5), (1, 2)],
mapname=maze,
encodingFunction=lambda x: ConsumableGridWorldIRL.stateVisitEncoding(
x, [(7, 5)]),
noise=noise,
binary=True)
#domain = Pinball(noise=0.3)
# Representation
representation = Tabular(domain, discretization=discretization)
# Policy
policy = eGreedy(representation, epsilon=0.3)
# Agent
opt["agent"] = Q_Learning(representation=representation,
policy=policy,
discount_factor=domain.discount_factor,
initial_learn_rate=0.1,
learn_rate_decay_mode="boyan",
boyan_N0=100,
lambda_=0.)
opt["checks_per_policy"] = 10
opt["max_steps"] = 150000
opt["num_policy_checks"] = 20
d = GoalPathPlanner(domain, representation, policy)
trajs = d.generateTrajectories(N=5)
dist = calculateStateDist((10, 7), trajs)
# Policy reset
policy = eGreedy(representation, epsilon=0.3)
representation = Tabular(domain, discretization=discretization)
opt["agent"] = Q_Learning(representation=representation,
policy=policy,
discount_factor=domain.discount_factor,
initial_learn_rate=0.1,
learn_rate_decay_mode="boyan",
boyan_N0=100,
lambda_=0.)
domain = ConsumableGridWorldIRL(
[(7, 5), (1, 2)],
mapname=maze,
encodingFunction=lambda x: ConsumableGridWorldIRL.stateVisitEncoding(
x, [(7, 5)]),
rewardFunction=lambda x, y, z, w: ConsumableGridWorldIRL.rewardIRL(
x, y, z, w, dist),
noise=noise)
pdomain = ConsumableGridWorldIRL(
[(7, 5), (1, 2)],
mapname=maze,
encodingFunction=lambda x: ConsumableGridWorldIRL.stateVisitEncoding(
x, [(7, 5)]),
noise=noise)
opt["domain"] = domain
experiment = Experiment(**opt)
experiment.run(visualize_steps=False,
performance_domain=pdomain,
visualize_learning=False,
visualize_performance=0)
experiment.save()
return np.max(experiment.result["return"]), np.sum(
experiment.result["return"])
def grid_world1_trp(exp_id=4, path="./Results/gridworld1"):
opt = {}
opt["exp_id"] = exp_id
opt["path"] = path
opt["checks_per_policy"] = 10
opt["max_steps"] = 150000
opt["num_policy_checks"] = 20
noise = 0.1
exp = 0.3
discretization = 20
# Domain:
maze = os.path.join(ConsumableGridWorld.default_map_dir,
'10x7-ACC2011.txt')
domain = ConsumableGridWorldIRL(
[(7, 5), (1, 2)],
mapname=maze,
encodingFunction=lambda x: ConsumableGridWorldIRL.stateVisitEncoding(
x, [(7, 5)]),
binary=True,
noise=noise)
#domain = Pinball(noise=0.3)
# Representation
representation = Tabular(domain, discretization=discretization)
# Policy
policy = eGreedy(representation, epsilon=0.3)
d = GoalPathPlanner(domain, representation, policy)
trajs = d.generateTrajectories(N=5)
a = TransitionStateClustering(window_size=2)
for t in trajs:
N = len(t)
demo = np.zeros((N, 2))
for i in range(0, N):
demo[i, :] = t[i][0:2]
a.addDemonstration(demo)
a.fit(normalize=False, pruning=0.5)
ac = [(round(a.means_[0][0]), round(a.means_[0][1])) for a in a.model]
print ac
#reinitialize
domain = ConsumableGridWorldIRL(
[(7, 5), (1, 2)],
mapname=maze,
encodingFunction=lambda x: ConsumableGridWorldIRL.statePassageEncoding(
x, ac, 5),
noise=noise)
representation = IncrementalTabular(domain, discretization=discretization)
policy = eGreedy(representation, epsilon=0.3)
opt["agent"] = Q_Learning(representation=representation,
policy=policy,
discount_factor=domain.discount_factor,
initial_learn_rate=0.1,
learn_rate_decay_mode="boyan",
boyan_N0=100,
lambda_=0.)
opt["domain"] = domain
experiment = Experiment(**opt)
experiment.run(visualize_steps=False,
visualize_learning=False,
visualize_performance=0)
experiment.save()
return np.max(experiment.result["return"]), np.sum(
experiment.result["return"])
def createPassageDomain(self,selfwaypoints,k=5):
return ConsumableGridWorldIRL(self.env_template["consumable"],
mapname=self.env_template["map"],
encodingFunction= lambda x: ConsumableGridWorldIRL.statePassageEncoding(x,waypoints,k),
noise=self.env_template["noise"],
binary=True)
def createSlidingDomain(self,k):
return ConsumableGridWorldIRL(self.env_template["consumable"],
mapname=self.env_template["map"],
encodingFunction= lambda x: ConsumableGridWorldIRL.slidingWindowEncoding(x,k),
noise=self.env_template["noise"])
def gridworld1_tirl(exp_id=7, path="./Results/gridworld1"):
opt = {}
opt["exp_id"] = exp_id
opt["path"] = path
opt["checks_per_policy"] = 10
opt["max_steps"] = 150000
opt["num_policy_checks"] = 20
noise = 0.1
exp = 0.3
discretization = 400
# Domain:
maze = os.path.join(ConsumableGridWorld.default_map_dir, '10x7-ACC2011.txt')
domain = ConsumableGridWorldIRL([(7,5), (1,2)],
mapname=maze,
encodingFunction= lambda x: ConsumableGridWorldIRL.stateVisitEncoding(x,[(7,5)]),
noise=noise,
binary=True)
#domain = Pinball(noise=0.3)
# Representation
representation = Tabular(domain, discretization=discretization)
# Policy
policy = eGreedy(representation, epsilon=0.3)
# Agent
opt["agent"] = Q_Learning(representation=representation, policy=policy,
discount_factor=domain.discount_factor,
initial_learn_rate=0.1,
learn_rate_decay_mode="boyan", boyan_N0=100,
lambda_=0.)
opt["checks_per_policy"] = 10
opt["max_steps"] = 150000
opt["num_policy_checks"] = 20
d = GoalPathPlanner(domain, representation,policy)
trajs = d.generateTrajectories(N=5)
dist = calculateStateDist((10,7), trajs)
a = TransitionStateClustering(window_size=2)
for t in trajs:
N = len(t)
demo = np.zeros((N,2))
for i in range(0,N):
demo[i,:] = t[i][0:2]
a.addDemonstration(demo)
a.fit(normalize=False, pruning=0.5)
dist = calculateStateDist((10,7), trajs)
ac = discrete2DClustersToPoints(a.model, dist, radius=1)
# Policy reset
policy = eGreedy(representation, epsilon=0.3)
representation = Tabular(domain, discretization=discretization)
opt["agent"] = Q_Learning(representation=representation, policy=policy,
discount_factor=domain.discount_factor,
initial_learn_rate=0.1,
learn_rate_decay_mode="boyan", boyan_N0=100,
lambda_=0.)
domain = ConsumableGridWorldIRL([(7,5), (1,2)],
mapname=maze,
encodingFunction= lambda x: ConsumableGridWorldIRL.stateVisitEncoding(x,[(7,5)]),
rewardFunction= lambda x,y,z,w: ConsumableGridWorldIRL.tRewardIRL(x,y,z,w,dist,[(7,5)]),
noise=noise)
pdomain = ConsumableGridWorldIRL([(7,5), (1,2)],
mapname=maze,
encodingFunction= lambda x: ConsumableGridWorldIRL.stateVisitEncoding(x,[(7,5)]),
noise=noise)
opt["domain"] = domain
experiment = Experiment(**opt)
experiment.run(visualize_steps=False,
performance_domain = pdomain,
visualize_learning=False,
visualize_performance=0)
experiment.save()
return np.max(experiment.result["return"]),np.sum(experiment.result["return"])
def gridworld1_rirl(exp_id=6, path="./Results/gridworld1"):
opt = {}
opt["exp_id"] = exp_id
opt["path"] = path
opt["checks_per_policy"] = 10
opt["max_steps"] = 150000
opt["num_policy_checks"] = 20
noise = 0.1
exp = 0.3
discretization = 400
# Domain:
maze = os.path.join(ConsumableGridWorld.default_map_dir, '10x7-ACC2011.txt')
domain = ConsumableGridWorldIRL([(7,5), (1,2)],
mapname=maze,
encodingFunction= lambda x: ConsumableGridWorldIRL.stateVisitEncoding(x,[(7,5)]),
noise=noise,
binary=True)
#domain = Pinball(noise=0.3)
# Representation
representation = Tabular(domain, discretization=discretization)
# Policy
policy = eGreedy(representation, epsilon=0.3)
# Agent
opt["agent"] = Q_Learning(representation=representation, policy=policy,
discount_factor=domain.discount_factor,
initial_learn_rate=0.1,
learn_rate_decay_mode="boyan", boyan_N0=100,
lambda_=0.)
opt["checks_per_policy"] = 10
opt["max_steps"] = 150000
opt["num_policy_checks"] = 20
d = GoalPathPlanner(domain, representation,policy)
trajs = d.generateTrajectories(N=5)
dist = calculateStateDist((10,7), trajs)
# Policy reset
policy = eGreedy(representation, epsilon=0.3)
representation = Tabular(domain, discretization=discretization)
opt["agent"] = Q_Learning(representation=representation, policy=policy,
discount_factor=domain.discount_factor,
initial_learn_rate=0.1,
learn_rate_decay_mode="boyan", boyan_N0=100,
lambda_=0.)
domain = ConsumableGridWorldIRL([(7,5), (1,2)],
mapname=maze,
encodingFunction= lambda x: ConsumableGridWorldIRL.stateVisitEncoding(x,[(7,5)]),
rewardFunction= lambda x,y,z,w: ConsumableGridWorldIRL.rewardIRL(x,y,z,w,dist),
noise=noise)
pdomain = ConsumableGridWorldIRL([(7,5), (1,2)],
mapname=maze,
encodingFunction= lambda x: ConsumableGridWorldIRL.stateVisitEncoding(x,[(7,5)]),
noise=noise)
opt["domain"] = domain
experiment = Experiment(**opt)
experiment.run(visualize_steps=False,
performance_domain = pdomain,
visualize_learning=False,
visualize_performance=0)
experiment.save()
return np.max(experiment.result["return"]),np.sum(experiment.result["return"])
