1
0
mirror of https://github.com/gryf/coach.git synced 2025-12-17 11:10:20 +01:00
Files
coach/rl_coach/agents/policy_optimization_agent.py
Gal Leibovich 49dea39d34 N-step returns for rainbow (#67)
* n_step returns for rainbow
* Rename CartPole_PPO -> CartPole_ClippedPPO
2018-11-07 18:33:08 +02:00

163 lines
6.9 KiB
Python

#
# Copyright (c) 2017 Intel Corporation
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
from collections import OrderedDict
from enum import Enum
from typing import Union
import numpy as np
from rl_coach.agents.agent import Agent
from rl_coach.core_types import Batch, ActionInfo
from rl_coach.logger import screen
from rl_coach.spaces import DiscreteActionSpace, BoxActionSpace
from rl_coach.utils import eps
class PolicyGradientRescaler(Enum):
TOTAL_RETURN = 0
FUTURE_RETURN = 1
FUTURE_RETURN_NORMALIZED_BY_EPISODE = 2
FUTURE_RETURN_NORMALIZED_BY_TIMESTEP = 3 # baselined
Q_VALUE = 4
A_VALUE = 5
TD_RESIDUAL = 6
DISCOUNTED_TD_RESIDUAL = 7
GAE = 8
## This is an abstract agent - there is no learn_from_batch method ##
class PolicyOptimizationAgent(Agent):
def __init__(self, agent_parameters, parent: Union['LevelManager', 'CompositeAgent']=None):
super().__init__(agent_parameters, parent)
self.policy_gradient_rescaler = None
if hasattr(self.ap.algorithm, 'policy_gradient_rescaler'):
self.policy_gradient_rescaler = self.ap.algorithm.policy_gradient_rescaler
# statistics for variance reduction
self.last_gradient_update_step_idx = 0
self.max_episode_length = 100000
self.mean_return_over_multiple_episodes = np.zeros(self.max_episode_length)
self.num_episodes_where_step_has_been_seen = np.zeros(self.max_episode_length)
self.entropy = self.register_signal('Entropy')
def log_to_screen(self):
# log to screen
log = OrderedDict()
log["Name"] = self.full_name_id
if self.task_id is not None:
log["Worker"] = self.task_id
log["Episode"] = self.current_episode
log["Total reward"] = round(self.total_reward_in_current_episode, 2)
log["Steps"] = self.total_steps_counter
log["Training iteration"] = self.training_iteration
screen.log_dict(log, prefix=self.phase.value)
def update_episode_statistics(self, episode):
episode_discounted_returns = []
for i in range(episode.length()):
transition = episode.get_transition(i)
episode_discounted_returns.append(transition.n_step_discounted_rewards)
self.num_episodes_where_step_has_been_seen[i] += 1
self.mean_return_over_multiple_episodes[i] -= self.mean_return_over_multiple_episodes[i] / \
self.num_episodes_where_step_has_been_seen[i]
self.mean_return_over_multiple_episodes[i] += transition.n_step_discounted_rewards / \
self.num_episodes_where_step_has_been_seen[i]
self.mean_discounted_return = np.mean(episode_discounted_returns)
self.std_discounted_return = np.std(episode_discounted_returns)
def train(self):
episode = self.current_episode_buffer
# check if we should calculate gradients or skip
num_steps_passed_since_last_update = episode.length() - self.last_gradient_update_step_idx
is_t_max_steps_passed = num_steps_passed_since_last_update >= self.ap.algorithm.num_steps_between_gradient_updates
if not (is_t_max_steps_passed or episode.is_complete):
return 0
total_loss = 0
if num_steps_passed_since_last_update > 0:
for network in self.networks.values():
network.set_is_training(True)
# we need to update the returns of the episode until now
episode.update_transitions_rewards_and_bootstrap_data()
# get t_max transitions or less if the we got to a terminal state
# will be used for both actor-critic and vanilla PG.
# In order to get full episodes, Vanilla PG will set the end_idx to a very big value.
transitions = episode[self.last_gradient_update_step_idx:]
batch = Batch(transitions)
# move the pointer for the last update step
if episode.is_complete:
self.last_gradient_update_step_idx = 0
else:
self.last_gradient_update_step_idx = episode.length()
# update the statistics for the variance reduction techniques
if self.policy_gradient_rescaler in \
[PolicyGradientRescaler.FUTURE_RETURN_NORMALIZED_BY_EPISODE,
PolicyGradientRescaler.FUTURE_RETURN_NORMALIZED_BY_TIMESTEP]:
self.update_episode_statistics(episode)
# accumulate the gradients
total_loss, losses, unclipped_grads = self.learn_from_batch(batch)
# apply the gradients once in every apply_gradients_every_x_episodes episodes
if self.current_episode % self.ap.algorithm.apply_gradients_every_x_episodes == 0:
for network in self.networks.values():
network.apply_gradients_and_sync_networks()
self.training_iteration += 1
for network in self.networks.values():
network.set_is_training(False)
# run additional commands after the training is done
self.post_training_commands()
return total_loss
def learn_from_batch(self, batch):
raise NotImplementedError("PolicyOptimizationAgent is an abstract agent. Not to be used directly.")
def get_prediction(self, states):
tf_input_state = self.prepare_batch_for_inference(states, "main")
return self.networks['main'].online_network.predict(tf_input_state)
def choose_action(self, curr_state):
# convert to batch so we can run it through the network
action_values = self.get_prediction(curr_state)
if isinstance(self.spaces.action, DiscreteActionSpace):
# DISCRETE
action_probabilities = np.array(action_values).squeeze()
action = self.exploration_policy.get_action(action_probabilities)
action_info = ActionInfo(action=action,
action_probability=action_probabilities[action])
self.entropy.add_sample(-np.sum(action_probabilities * np.log(action_probabilities + eps)))
elif isinstance(self.spaces.action, BoxActionSpace):
# CONTINUOUS
action = self.exploration_policy.get_action(action_values)
action_info = ActionInfo(action=action)
else:
raise ValueError("The action space of the environment is not compatible with the algorithm")
return action_info