def test_discrete_gridworld_experiment():
"""Test the learning on an n*n tabular, discrete gridworld."""
environment = Gridworld(3, 3, 0.0)
experiment_description = ExperimentDescriptor(5, 100)
agent = Agent(RandomPolicy(environment.num_actions), TabularFA(environment.num_states(), environment.num_actions),
environment.num_actions, 0.1, 0.99, 0.1, 0.1, 0.05, 10, 0, 0)
results_path = os.path.join(os.getcwd(), 'results.txt')
results_descriptor = ResultsDescriptor(100, results_path, ['interval_id', 'steps'])
start(experiment_description, agent, environment, results_descriptor)
def test_discrete_gridworld_experiment():
"""Test the learning on an n*n tabular, discrete gridworld."""
environment = Gridworld(3, 3, 0.0)
experiment_description = ExperimentDescriptor(5, 100)
agent = Agent(RandomPolicy(environment.num_actions),
TabularFA(environment.num_states(), environment.num_actions),
environment.num_actions, 0.1, 0.99, 0.1, 0.1, 0.05, 10, 0, 0)
results_path = os.path.join(os.getcwd(), 'results.txt')
results_descriptor = ResultsDescriptor(100, results_path,
['interval_id', 'steps'])
start(experiment_description, agent, environment, results_descriptor)
def main(argv):
"""Execute experiment."""
args = parse_args(argv)
random.seed(args.seed)
logging.basicConfig(level=log_level(args.log_level))
environment = (args.environment == 'gridworld' and Gridworld(args.gridworld_width, args.gridworld_height, args.failure_rate)) or \
(args.environment == 'gridworld_continuous' and GridworldContinuous(0.2, 0.01)) or \
(args.environment == 'combo_lock' and CombinationLock(args.gridworld_height, args.gridworld_width, 4, args.failure_rate))
policy = (args.agent_policy == 'random'
and RandomPolicy(environment.num_actions))
fa = ((args.environment == 'gridworld' or args.environment == 'combo_lock') and
TabularFA(environment.num_states(), environment.num_actions)) or \
(args.environment == 'gridworld_continuous' and RBF(2, 5, environment.num_actions, beta=args.beta))
agent = Agent(policy,
fa,
environment.num_actions,
args.alpha,
args.gamma,
args.eta,
args.zeta,
args.epsilon,
args.num_vi,
args.sim_samples,
args.sim_steps,
retain_theta=args.retain_theta,
subgoals=environment.create_subgoals())
agent.policy = RandomOptionPolicy(agent, args.random_options)
if args.agent_viz:
agent.create_visualization(
args.environment == 'gridworld'
or args.environment == 'combo_lock', environment)
# agent.exploit(np.asarray([1, 1]))
# agent.exploit(environment.num_states()-1)
# results_descriptor = ResultsDescriptor(args.results_interval, args.results_path, ['interval_id', 'steps'])
# experiment_descriptor = ExperimentDescriptor(args.plan_interval, args.num_steps)
# start(experiment_descriptor, agent, environment, results_descriptor)
e = Experiment2(agent, environment, args.plan_interval, args.num_steps,
args.viz_steps)
e.run()
def test_termination():
"""Does the environment terminate in the correct state?"""
env = Gridworld(4, 4, 0)
assert not env.is_terminal(env.state_from_grid_position(GridPosition(0, 0)))
assert env.is_terminal(env.state_from_grid_position(GridPosition(3, 3)))
def test_taking_actions():
"""Does the environment correctly change the state when told to take an action with and without stochasticity?"""
random.seed()
env = Gridworld(4, 4, 0.0)
# Deterministic tests
assert env.next_state(env.initial_state(), Action.up) == env.state_from_grid_position(GridPosition(0, 1))
assert env.next_state(env.initial_state(), Action.down) == env.state_from_grid_position(GridPosition(0, 0))
assert env.next_state(env.initial_state(), Action.left) == env.state_from_grid_position(GridPosition(0, 0))
assert env.next_state(env.initial_state(), Action.right) == env.state_from_grid_position(GridPosition(1, 0))
# Stochastic tests
env.failure_rate = 0.1
assert ratio_test(lambda state: state == env.state_from_grid_position(GridPosition(0, 0)), partial(env.next_state, env.initial_state(), Action.left), 10000) == 1.0
ratio = ratio_test(lambda state: state == env.state_from_grid_position(GridPosition(0, 0)), partial(env.next_state, env.initial_state(), Action.up), 10000)
assert 0.09 < ratio < 0.11
