def test_tabular_function():
def reward_function_factory(env):
params = np.zeros(64)
params[-1] = 1.
return TabularRewardFunction(env, params)
env = make_wrapped_env('FrozenLake8x8-v0',
reward_function_factory=reward_function_factory,
with_model_wrapper=True)
agent = ValueIteration(env)
agent.train(1)
trajs = collect_trajs(env, agent, 10)
for traj in trajs:
for i in range(len(traj['rewards'])):
assert np.isclose(traj['rewards'][i], traj['true_rewards'][i])
def evaluate(self, estim_rewards, ord=2):
# value function for actual policy
expert_agent = ValueIteration(env=self.env, gamma=self.gamma)
expert_agent.train(time_limit=50)
value_actual = expert_agent.V
# value function for learned policy
learned_agent = ValueIteration(env=self.env, gamma=self.gamma)
learned_agent.train(time_limit=50, reward_function=estim_rewards)
value_learned = learned_agent.V
# Inverse Learning Error
ile = np.linalg.norm((value_actual - value_learned), ord=ord)
return ile
def rl_alg_factory(env):
return ValueIteration(env, {'gamma': 0.9})
def rl_alg_factory(env):
return ValueIteration(env)
from irl_benchmark.irl.reward.reward_wrapper import RewardWrapper
from irl_benchmark.rl.algorithms.value_iteration import ValueIteration
from irl_benchmark.utils.utils import get_transition_matrix
store_to = 'data/frozen/expert/'
no_episodes = 1000
max_steps_per_episode = 1000
env = gym.make('FrozenLake8x8-v0')
env = FrozenLakeFeatureWrapper(env)
initial_reward_function_estimate = FeatureBasedRewardFunction(
env=env, parameters=np.zeros(64))
env = RewardWrapper(env=env, reward_function=initial_reward_function_estimate)
# Generate expert trajectories.
expert_agent = ValueIteration(env)
print('Training expert agent...')
expert_agent.train(30)
expert_trajs = collect_trajs(env, expert_agent, no_episodes,
max_steps_per_episode, store_to)
feat_map = np.eye(64)
transition_dynamics = get_transition_matrix(env)
def rl_alg_factory(env):
return ValueIteration(env)
meirl = MaxEnt(env,
def rl_alg_factory(env, lp=False):
'''Return an RL algorithm that will collect expert trajectories.'''
if lp:
return ValueIteration(env, error=1e-5)
else:
return TabularQ(env)