def evaluate_policy(policy, env, eval_episodes=10):
reward_arr = np.zeros(eval_episodes)
for i in range(eval_episodes):
obs = env.reset()
done = False
total_reward = 0.
while not done:
feasible_actions = AllocationEnv.get_feasible_actions(
obs["board_config"])
action_mask = AllocationEnv.get_action_mask(
feasible_actions, env.action_space.n)
action, _states = policy.predict(obs, mask=action_mask)
action = AllocationEnv.check_action(obs['board_config'], action)
obs, reward, done, _ = env.step(action)
total_reward += reward
reward_arr[i] = total_reward
avg_reward = reward_arr.mean()
std_reward = reward_arr.std()
print("---------------------------------------")
print("Evaluation over {} episodes: {:.1f} ({:.2f})".format(
eval_episodes, avg_reward, std_reward))
print("---------------------------------------")
return avg_reward, std_reward
def evaluate(self):
gamma = .8
rewards = []
for i in range(self.n_episodes):
self.queue = self.build_queue(self.buffer)
r_i = 0
state, _, _, _, _ = self.buffer.sample(batch_size=1)
state = state[0]
iter = 0
cntr = 0
while True:
board_cfg = State.get_board_config_from_vec(
state,
n_regions=self.n_regions,
n_products=self.n_products)
feasible_actions = AllocationEnv.get_feasible_actions(
board_cfg)
action_mask = AllocationEnv.get_action_mask(
feasible_actions, self.n_actions)
#M = self.get_m(state, action_mask)
M = 1
try:
_, a, r, s_prime = self.queue[state].pop()
#_, a, r, s_prime = self.queue[state][-1]
except IndexError:
break
alpha = random.random()
prob_policy = self.policy.proba_step(state.reshape(1, -1),
mask=action_mask)[0][a]
prob_env = self.env_policy.predict_proba(state)[a]
rejection_tol = (1 / M) * prob_policy / prob_env
iter += 1
print(f"eps: {i+1} - iter:{iter} - success: {cntr}")
if alpha > rejection_tol:
continue
else:
#self.queue[state].pop()
r_i += (gamma)**cntr * r
state = s_prime
cntr += 1
if r_i > 0:
rewards.append(r_i)
return rewards
def learn(self):
for i in range(self.epochs):
print(f"Epoch {i}/{self.epochs}")
pbar = tqdm(range(self.rollout_batch_size))
for b in pbar:
#state = self.buffer_env.sample(batch_size=1)[0][0]
state = self.env_model.reset()
state = State.get_vec_observation(state)
for h in range(self.rollout):
pbar.set_description(f"batch: {b} rollout: {h}")
board_cfg = State.get_board_config_from_vec(state,
n_regions=self.n_regions,
n_products=self.n_products
)
feasible_actions = AllocationEnv.get_feasible_actions(board_cfg)
#feasible_actions = AllocationEnv.get_feasible_actions(state["board_config"])
action_mask = AllocationEnv.get_action_mask(feasible_actions, self.n_actions)
# sample action a_j ~ pi(s_j)
alpha = random.random()
if alpha < self.eps:
action = self.env_model.action_space.sample()
else:
action, _states = self.policy.predict(state.reshape(1, -1), mask=action_mask)
# compute dynamics from env model
new_state, r_hat, dones, info = self.env_model.step(action)
new_state = State.get_vec_observation(new_state)
reward = self.get_penalized_reward(r_hat, self.lmbda)
# add (s, a, r, s') to buffer
self.buffer_model.add(obs_t=state,
action=action,
reward=reward,
obs_tp1=new_state,
done=float(dones))
state = new_state
# update policy with samples from D_env and D_model
self.policy.update_weights(self.buffer_model)
self.save_buffer()
def map_optimal_rewards():
state = env.reset()
total_reward = 0
results = {'rewards': [0.0]}
optimal_actions = []
curr_action = 0
for day in range(TEST_T):
curr_state = copy.deepcopy(env.state)
feasible_actions = AllocationEnv.get_feasible_actions(
curr_state.board_config)
proposed_action = np.random.choice(list(feasible_actions))
curr_state_step, curr_reward, b, i = env.step(curr_action)
env.set_state(curr_state)
proposed_state, proposed_reward, b, i = env.step(proposed_action)
curr_f = get_f(ae=-curr_reward, lmbda=LMBDA, log=True, T=T)
proposed_f = get_f(ae=-proposed_reward, lmbda=LMBDA, log=True, T=T)
gamma = get_gamma(f_current=curr_f, f_proposed=proposed_f, log=True)
# Generate random number on log scale
sr = np.log(random.random())
if sr < gamma: # made progress
#state, final_reward, _, _ = env.step(curr_best_action) # update the state after each day based on the optimal action taken
optimal_actions.append(proposed_action)
curr_best_action = proposed_action
final_reward = proposed_reward
else:
optimal_actions.append(curr_action)
state, final_reward, _, _ = env.step(curr_action)
curr_best_action = curr_action
total_reward += final_reward
results['rewards'].append(total_reward)
print("best action: {} - reward: {}".format(curr_best_action,
final_reward))
print("total reward: {}".format(total_reward))
return state, optimal_actions, results
def map_optimal_rewards(tabu_len, k):
state = env.reset()
total_reward = 0
results = {'rewards': [0.0]}
optimal_actions = []
for day in range(TEST_T):
curr_best_val = 0.0
curr_best_action = 0.0
curr_state = copy.deepcopy(env.state)
feasible_actions = AllocationEnv.get_feasible_actions(curr_state.board_config)
for action in feasible_actions:
print("Iteration: {}, Action: {}".format(day, action), end='\r')
action = AllocationEnv.check_action(curr_state.board_config, action)
proposed_state, reward, b, i = env.step(action)
env.set_state(curr_state)
if reward > curr_best_val:
curr_best_action = action
optimal_actions.append(curr_best_action)
curr_best_action = AllocationEnv.check_action(curr_state.board_config, curr_best_action)
state, final_reward, _ , _ = env.step(curr_best_action) # update the state after each day based on the optimal action taken
total_reward += final_reward
curr_best_val = final_reward
results['rewards'].append(total_reward)
print("best action: {} - reward: {}".format(curr_best_action, final_reward))
print("total reward: {}".format(total_reward))
return state, optimal_actions, results
exploration_fraction=0.35,
exploration_final_eps=0.2)
model.learn(total_timesteps=TIME_STEPS, learning_curve=False, test_t=TEST_T)
with open(f"../data/{store_id}-buffer-d-test.p", 'wb') as f:
pickle.dump(model.replay_buffer, f)
results = {'rewards': [0.0]}
buffer = ReplayBuffer(size=50000)
for j in range(100):
obs = env.reset()
for i in range(TEST_T):
feasible_actions = AllocationEnv.get_feasible_actions(
obs["board_config"])
action_mask = AllocationEnv.get_action_mask(feasible_actions,
n_actions)
action, _states = model.predict(obs, mask=action_mask)
action = AllocationEnv.check_action(obs['board_config'], action)
new_obs, r, dones, info = env.step([action])
results['rewards'].append(r[0] + results['rewards'][-1])
# add (s, a, r, s') to buffer
buffer.add(obs_t=State.get_vec_observation(obs),
action=action,
reward=r[0],
obs_tp1=State.get_vec_observation(new_obs),
done=float(dones))
def learn(self,
total_timesteps,
callback=None,
seed=None,
log_interval=100,
tb_log_name="DQN",
reset_num_timesteps=True,
replay_wrapper=None,
learning_curve=False,
test_t=None):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
self._setup_learn(seed)
# Create the replay buffer
if self.prioritized_replay:
self.replay_buffer = PrioritizedReplayBuffer(
self.buffer_size, alpha=self.prioritized_replay_alpha)
if self.prioritized_replay_beta_iters is None:
prioritized_replay_beta_iters = total_timesteps
else:
prioritized_replay_beta_iters = self.prioritized_replay_beta_iters
self.beta_schedule = LinearSchedule(
prioritized_replay_beta_iters,
initial_p=self.prioritized_replay_beta0,
final_p=1.0)
else:
self.replay_buffer = ReplayBuffer(self.buffer_size)
self.beta_schedule = None
if replay_wrapper is not None:
assert not self.prioritized_replay, "Prioritized replay buffer is not supported by HER"
self.replay_buffer = replay_wrapper(self.replay_buffer)
# Create the schedule for exploration starting from 1.
self.exploration = LinearSchedule(
schedule_timesteps=int(self.exploration_fraction *
total_timesteps),
initial_p=1.0,
final_p=self.exploration_final_eps)
episode_rewards = [0.0]
self.cumul_reward = [0.0]
episode_successes = []
obs = self.env.reset()
reset = True
self.episode_reward = np.zeros((1, ))
# variables for test eval ##
test_step = test_t * 3
test_results = {'sum': []}
test_ts = []
for _ in range(total_timesteps):
## Test eval period ##
if learning_curve and _ % test_step == 0 and _ > 0:
print("--> Simulating test period")
self.env.reset()
test_r = 0.0
for i in range(test_t):
feasible_actions = AllocationEnv.get_feasible_actions(
obs["board_config"])
action_mask = AllocationEnv.get_action_mask(
feasible_actions, self.env.action_space.n)
action, _states = self.predict(obs, mask=action_mask)
action = AllocationEnv.check_action(
obs['board_config'], action)
obs, rewards, dones, info = self.env.step(action)
test_r += rewards
test_results["sum"].append(test_r)
test_ts.append(_)
self.env.reset()
# plot test eval progress
plt.plot(test_ts, test_results["sum"])
# plt.errorbar(iteration_cuts, results["mean"], yerr=results["std"], fmt='.k')
plt.xlabel("Iteration count")
plt.ylabel("Total (sum) test reward")
plt.savefig("figs/rl-learning-curve-{}.pdf".format(
cfg.vals['prj_name']))
plt.clf()
plt.close()
# write test eval progress
write_results = {}
for k, v in test_results.items():
write_results[k] = serialize_floats(v)
with open(
"output/rl-learning-curve-{}.json".format(
cfg.vals['prj_name']), 'w') as f:
json.dump(write_results, f)
if callback is not None:
# Only stop training if return value is False, not when it is None. This is for backwards
# compatibility with callbacks that have no return statement.
if callback(locals(), globals()) is False:
break
# Take action and update exploration to the newest value
kwargs = {}
if not self.param_noise:
update_eps = self.exploration.value(self.num_timesteps)
update_param_noise_threshold = 0.
else:
update_eps = 0.
# Compute the threshold such that the KL divergence between perturbed and non-perturbed
# policy is comparable to eps-greedy exploration with eps = exploration.value(t).
# See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017
# for detailed explanation.
update_param_noise_threshold = \
-np.log(1. - self.exploration.value(self.num_timesteps) +
self.exploration.value(self.num_timesteps) / float(self.env.action_space.n))
kwargs['reset'] = reset
kwargs[
'update_param_noise_threshold'] = update_param_noise_threshold
kwargs['update_param_noise_scale'] = True
feasible_actions = AllocationEnv.get_feasible_actions(
obs["board_config"])
action_mask = AllocationEnv.get_action_mask(
feasible_actions, self.action_space.n)
with self.sess.as_default():
action = self.act(State.get_vec_observation(obs)[None],
update_eps=update_eps,
**kwargs,
mask=action_mask)[0]
reset = False
# CHECK IF ACTIONS IS FEASIBLE
action = AllocationEnv.check_action(obs['board_config'],
action)
env_action = action
new_obs, rew, done, info = self.env.step(env_action)
print("action: {} - reward: {} - eps: {:.4}".format(
action, rew, update_eps))
print(new_obs['day_vec'])
print(new_obs['board_config'])
# Store transition in the replay buffer.
self.replay_buffer.add(State.get_vec_observation(obs), action,
rew, State.get_vec_observation(new_obs),
float(done))
obs = new_obs
if writer is not None:
ep_rew = np.array([rew]).reshape((1, -1))
ep_done = np.array([done]).reshape((1, -1))
self.episode_reward = total_episode_reward_logger(
self.episode_reward, ep_rew, ep_done, writer,
self.num_timesteps)
episode_rewards[-1] += rew
self.cumul_reward.append(self.cumul_reward[-1] + rew)
if done:
maybe_is_success = info.get('is_success')
if maybe_is_success is not None:
episode_successes.append(float(maybe_is_success))
if not isinstance(self.env, VecEnv):
obs = self.env.reset()
episode_rewards.append(0.0)
reset = True
# Do not train if the warmup phase is not over
# or if there are not enough samples in the replay buffer
can_sample = self.replay_buffer.can_sample(self.batch_size)
if can_sample and self.num_timesteps > self.learning_starts \
and self.num_timesteps % self.train_freq == 0:
# Minimize the error in Bellman's equation on a batch sampled from replay buffer.
if self.prioritized_replay:
experience = self.replay_buffer.sample(
self.batch_size,
beta=self.beta_schedule.value(self.num_timesteps))
(obses_t, actions, rewards, obses_tp1, dones, weights,
batch_idxes) = experience
else:
obses_t, actions, rewards, obses_tp1, dones = self.replay_buffer.sample(
self.batch_size)
weights, batch_idxes = np.ones_like(rewards), None
if writer is not None:
# run loss backprop with summary, but once every 100 steps save the metadata
# (memory, compute time, ...)
if (1 + self.num_timesteps) % 100 == 0:
run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
summary, td_errors = self._train_step(
obses_t,
actions,
rewards,
obses_tp1,
obses_tp1,
dones,
weights,
sess=self.sess,
options=run_options,
run_metadata=run_metadata)
writer.add_run_metadata(
run_metadata, 'step%d' % self.num_timesteps)
else:
summary, td_errors = self._train_step(
obses_t,
actions,
rewards,
obses_tp1,
obses_tp1,
dones,
weights,
sess=self.sess)
writer.add_summary(summary, self.num_timesteps)
else:
_, td_errors = self._train_step(obses_t,
actions,
rewards,
obses_tp1,
obses_tp1,
dones,
weights,
sess=self.sess)
if self.prioritized_replay:
new_priorities = np.abs(
td_errors) + self.prioritized_replay_eps
self.replay_buffer.update_priorities(
batch_idxes, new_priorities)
if can_sample and self.num_timesteps > self.learning_starts and \
self.num_timesteps % self.target_network_update_freq == 0:
# Update target network periodically.
self.update_target(sess=self.sess)
if len(episode_rewards[-101:-1]) == 0:
mean_100ep_reward = -np.inf
else:
mean_100ep_reward = round(
float(np.mean(episode_rewards[-101:-1])), 1)
num_episodes = len(episode_rewards)
if self.verbose >= 1 and done and log_interval is not None and len(
episode_rewards) % log_interval == 0:
logger.record_tabular("steps", self.num_timesteps)
logger.record_tabular("episodes", num_episodes)
if len(episode_successes) > 0:
logger.logkv("success rate",
np.mean(episode_successes[-100:]))
logger.record_tabular("mean 100 episode reward",
mean_100ep_reward)
logger.record_tabular(
"% time spent exploring",
int(100 * self.exploration.value(self.num_timesteps)))
logger.dump_tabular()
print('timestamp: {}'.format(self.num_timesteps, end='\r\n'))
self.num_timesteps += 1
return self
