rewards[-1] = rewards[-1] * 10
dataset.episodes.append(Episode(states, actions, rewards, pb_sas))
return dataset
if __name__ == '__main__':
# np.random.seed(RANDOM_SEED)
s = env.reset()
n_states = poly.fit_transform(s.reshape(1, -1))
agent = Agent(n_states.shape[1],
env.action_space.n,
delta=0.2,
sigma=0.01,
is_tabular=False)
agent.c = -np.inf
mean_return = 0
did_improve = []
safety_dataset = generate_dataset(env, agent, 1000)
candidate_dataset = generate_dataset(env, agent, 1000)
ngen = 1
for epoch in range(1000):
print(f'Epoch: {epoch}')
print('---------------')
did_pass = agent.update(safety_dataset,
candidate_dataset,
1,
write=False)
if did_pass:
actions.append(a)
rewards.append(r)
pb_sas.append(action_probs[a])
s = poly.fit_transform(next_state.reshape(1, -1))
rewards[-1] = rewards[-1] * 10
dataset.episodes.append(Episode(states, actions, rewards, pb_sas))
return dataset
if __name__ == '__main__':
# np.random.seed(RANDOM_SEED)
s = env.reset()
n_states = poly.fit_transform(s.reshape(1, -1))
agent = Agent(n_states.shape[1], env.action_space.n, delta=0.1, sigma=0.1, is_tabular=False)
agent.c = 0
mean_return = 0
did_improve = []
safety_dataset = generate_dataset(env, agent, 1000)
candidate_dataset = generate_dataset(env, agent, 1000)
ngen = 1
for epoch in range(1000):
print(f'Epoch: {epoch}')
print('---------------')
did_pass = agent.update(safety_dataset, candidate_dataset, 1, write=False)
if did_pass:
eval_dataset = generate_dataset(env, agent, 1000)
gt_estimates = agent.expected_discounted_return(eval_dataset)
next_mean_return = np.mean(gt_estimates)
else:
success.append(1)
rewards[-1] = rewards[-1] * 10
dataset.episodes.append(Episode(states, actions, rewards, pb_sas))
print(f'Success rate: {np.mean(success)}')
return dataset
if __name__ == '__main__':
# np.random.seed(RANDOM_SEED)
agent = Agent(env.observation_space.n,
env.action_space.n,
delta=0.25,
sigma=0.01)
agent.c = -10
mean_return = 0
did_improve = []
safety_dataset = generate_dataset(env, agent, 5000)
candidate_dataset = generate_dataset(env, agent, 5000)
ngen = 1
for epoch in range(1000):
print(f'Epoch: {epoch}')
print('---------------')
did_pass = agent.update(safety_dataset,
candidate_dataset,
1,
write=False)
if did_pass: