def test_MonteCarloFirstVisit():
env = rl.Taxi()
agent = rl.backend.MonteCarloFirstVisit(states=env.states,
actions=env.actions,
epsilon=0.03,
gamma=0.99)
session = rl.Session(env, agent)
session.run(100, 10, 10)
def test_Double_Expected_Sarsa():
env = rl.Taxi()
agent = rl.DoubleExpectedSarsa(states=env.states,
actions=env.actions,
alpha=0.1,
epsilon=0.03,
gamma=0.99)
session = rl.Session(env, agent)
session.run(100, 10, 10)
def test_MonteCarlo():
env = rl.Taxi()
agent = rl.MonteCarlo(states=env.states,
actions=env.actions,
epsilon=0.03,
gamma=0.99,
visit_update='first')
session = rl.Session(env, agent)
session.run(100, 10, 10)
agent = rl.MonteCarlo(states=env.states,
actions=env.actions,
epsilon=0.03,
gamma=0.99,
visit_update='every')
session = rl.Session(env, agent)
session.run(100, 10, 10)
def test_Double_QLearning():
env = rl.Taxi()
agent = rl.DoubleQLearning(states=env.states,
actions=env.actions,
alpha=0.1,
epsilon=0.03,
gamma=0.99)
session = rl.Session(env, agent)
session.run(100, 10, 10)
def test_DoubleSarsa():
env = rl.Taxi()
agent = rl.DoubleSarsa(states=env.states,
actions=env.actions,
alpha=0.1,
epsilon=0.03,
gamma=0.99,
backend='cpp')
session = rl.Session(env, agent)
session.run(100, 10, 10)
def test_n_step_sarsa():
env = rl.Taxi()
agent = rl.NStepSarsa(states=env.states,
actions=env.actions,
alpha=0.1,
epsilon=0.03,
gamma=0.99,
n_step=5)
session = rl.Session(env, agent)
session.run(100, 10, 10)
def test_Sarsa():
env = rl.Taxi()
agent = rl.backend.Sarsa(
states=env.states,
actions=env.actions,
alpha=0.1,
epsilon=0.03,
gamma=0.99
)
session = rl.Session(env, agent)
session.run(100, 10, 10)
def test_TabularDynaQPlus():
env = rl.Taxi()
agent = rl.TabularDynaQPlus(states=env.states,
actions=env.actions,
alpha=0.1,
epsilon=0.03,
gamma=0.99,
n_plan=5,
backend='cpp')
session = rl.Session(env, agent)
session.run(100, 10, 10)
def test_NStepTreeBackup():
env = rl.Taxi()
agent = rl.NStepTreeBackup(states=env.states,
actions=env.actions,
alpha=0.1,
epsilon=0.03,
gamma=0.99,
n_step=5,
backend='cpp')
session = rl.Session(env, agent)
session.run(100, 10, 10)
def test_sarsa_appr():
env = rl.Taxi()
agent = rl.SemiGradientSarsa(actions=env.actions,
alpha=0.1,
epsilon=0.03,
gamma=0.99,
features=1,
tilings=2,
tilings_offset=[1],
tile_size=[1])
session = rl.Session(env, agent)
session.run(100, 10, 10)
def test_PrioritizedSweeping():
env = rl.Taxi()
agent = rl.PrioritizedSweeping(states=env.states,
actions=env.actions,
alpha=0.1,
epsilon=0.03,
gamma=0.99,
n_plan=5,
theta=0.5,
backend='cpp')
session = rl.Session(env, agent)
session.run(100, 10, 10)
def test_SemiGradientExpectedSarsa():
env = rl.Taxi()
agent = rl.SemiGradientExpectedSarsa(actions=env.actions,
alpha=0.1,
epsilon=0.01,
gamma=0.99,
features=1,
tilings=2,
tilings_offset=[4, 4],
tile_size=[4, 4],
backend='cpp')
session = rl.Session(env, agent)
session.run(100, 10, 10)
def test_semigradient_n_step_expected_sarsa():
env = rl.Taxi()
agent = rl.SemiGradientNStepExpectedSarsa(actions=env.actions,
alpha=0.1,
epsilon=0.03,
gamma=0.99,
n_step=5,
features=1,
tilings=2,
tilings_offset=[1],
tile_size=[1])
session = rl.Session(env, agent)
session.run(100, 10, 10)
def test_SemiGradientMonteCarlo():
env = rl.Taxi()
agent = rl.backend.SemiGradientMonteCarlo(
actions=env.actions,
alpha=0.1,
epsilon=0.01,
gamma=0.99,
features=1,
tilings=2,
tilings_offset=[4, 4],
tile_size=[4, 4]
)
session = rl.Session(env, agent)
session.run(100, 10, 10)
def test_sarsa_lambda():
env = rl.Taxi()
agent = rl.SarsaLambda(
actions=env.actions,
alpha=0.1,
epsilon=0.03,
gamma=0.99,
lambd=0.98,
trace_type="replacing",
features=1,
tilings=4,
)
session = rl.Session(env, agent)
session.run(100, 10, 10)
observe_confounder=observe_confounder, build_causal_model=True)
elif env_type == 'confounder_not_directly_influencing_outcome':
env = rl.ConfounderNotDirectlyInfluencingOutcome(
observe_confounder=observe_confounder, build_causal_model=True)
results = []
####################################
# Random Agent
####################################
random_agents = [
rl.Random(actions=env.actions) for _ in range(number_of_agents)
]
for agent in random_agents:
session = rl.Session(env=env, agent=agent, max_steps=max_steps)
results.append(
session.run(episodes=episodes,
test_offset=test_offset,
test_samples=test_sample,
render=False))
####################################
# A lot of Causal Q-Learning
####################################
causal_q_learning_agents = [
rl.CausalQLearning(states=env.states,
actions=env.actions,
alpha=alpha,
epsilon=epsilon,
epsilon_decay=epsilon_decay,
env = rl.TaxiGenericModel(build_causal_model=True,
observability='partial') # full, partial
elif env_name == 'taxi_confounder':
env = rl.TaxiConfounder(build_causal_model=True,
observe_confounder=observe_confounders)
results = []
####################################
# Random agent
####################################
random_agents = [
rl.Random(actions=env.actions) for _ in range(number_of_agents)
]
for agent in random_agents:
session = rl.Session(env=env, agent=agent)
results.append(
session.run(episodes=episodes,
test_offset=test_offset,
test_samples=test_sample,
render=False))
####################################
# Causal Q-Learning
####################################
causal_q_learning_agents = [
rl.CausalQLearning(states=env.states,
actions=env.actions,
alpha=alpha,
epsilon=epsilon,
epsilon_decay=epsilon_decay,
