ppo_agent.py

#
# 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.
#

import copy
from collections import OrderedDict
from typing import Union

import numpy as np

from rl_coach.agents.actor_critic_agent import ActorCriticAgent
from rl_coach.agents.policy_optimization_agent import PolicyGradientRescaler
from rl_coach.architectures.embedder_parameters import InputEmbedderParameters
from rl_coach.architectures.head_parameters import PPOHeadParameters, VHeadParameters
from rl_coach.architectures.middleware_parameters import FCMiddlewareParameters
from rl_coach.base_parameters import AlgorithmParameters, NetworkParameters, \
    AgentParameters, DistributedTaskParameters

from rl_coach.core_types import EnvironmentSteps, Batch
from rl_coach.exploration_policies.additive_noise import AdditiveNoiseParameters
from rl_coach.exploration_policies.categorical import CategoricalParameters
from rl_coach.logger import screen
from rl_coach.memories.episodic.episodic_experience_replay import EpisodicExperienceReplayParameters
from rl_coach.spaces import DiscreteActionSpace, BoxActionSpace
from rl_coach.utils import force_list


class PPOCriticNetworkParameters(NetworkParameters):
    def __init__(self):
        super().__init__()
        self.input_embedders_parameters = {'observation': InputEmbedderParameters(activation_function='tanh')}
        self.middleware_parameters = FCMiddlewareParameters(activation_function='tanh')
        self.heads_parameters = [VHeadParameters()]
        self.async_training = True
        self.l2_regularization = 0
        self.create_target_network = True
        self.batch_size = 128


class PPOActorNetworkParameters(NetworkParameters):
    def __init__(self):
        super().__init__()
        self.input_embedders_parameters = {'observation': InputEmbedderParameters(activation_function='tanh')}
        self.middleware_parameters = FCMiddlewareParameters(activation_function='tanh')
        self.heads_parameters = [PPOHeadParameters()]
        self.optimizer_type = 'Adam'
        self.async_training = True
        self.l2_regularization = 0
        self.create_target_network = True
        self.batch_size = 128


class PPOAlgorithmParameters(AlgorithmParameters):
    """
    :param policy_gradient_rescaler: (PolicyGradientRescaler)
        This represents how the critic will be used to update the actor. The critic value function is typically used
        to rescale the gradients calculated by the actor. There are several ways for doing this, such as using the
        advantage of the action, or the generalized advantage estimation (GAE) value.

    :param gae_lambda: (float)
        The :math:`\lambda` value is used within the GAE function in order to weight different bootstrap length
        estimations. Typical values are in the range 0.9-1, and define an exponential decay over the different
        n-step estimations.

    :param target_kl_divergence: (float)
        The target kl divergence between the current policy distribution and the new policy. PPO uses a heuristic to
        bring the KL divergence to this value, by adding a penalty if the kl divergence is higher.

    :param initial_kl_coefficient: (float)
        The initial weight that will be given to the KL divergence between the current and the new policy in the
        regularization factor.

    :param high_kl_penalty_coefficient: (float)
        The penalty that will be given for KL divergence values which are highes than what was defined as the target.

    :param clip_likelihood_ratio_using_epsilon: (float)
        If not None, the likelihood ratio between the current and new policy in the PPO loss function will be
        clipped to the range [1-clip_likelihood_ratio_using_epsilon, 1+clip_likelihood_ratio_using_epsilon].
        This is typically used in the Clipped PPO version of PPO, and should be set to None in regular PPO
        implementations.

    :param value_targets_mix_fraction: (float)
        The targets for the value network are an exponential weighted moving average which uses this mix fraction to
        define how much of the new targets will be taken into account when calculating the loss.
        This value should be set to the range (0,1], where 1 means that only the new targets will be taken into account.

    :param estimate_state_value_using_gae: (bool)
        If set to True, the state value will be estimated using the GAE technique.

    :param use_kl_regularization: (bool)
        If set to True, the loss function will be regularized using the KL diveregence between the current and new
        policy, to bound the change of the policy during the network update.

    :param beta_entropy: (float)
        An entropy regulaization term can be added to the loss function in order to control exploration. This term
        is weighted using the :math:`\beta` value defined by beta_entropy.

    """
    def __init__(self):
        super().__init__()
        self.policy_gradient_rescaler = PolicyGradientRescaler.GAE
        self.gae_lambda = 0.96
        self.target_kl_divergence = 0.01
        self.initial_kl_coefficient = 1.0
        self.high_kl_penalty_coefficient = 1000
        self.clip_likelihood_ratio_using_epsilon = None
        self.value_targets_mix_fraction = 0.1
        self.estimate_state_value_using_gae = True
        self.use_kl_regularization = True
        self.beta_entropy = 0.01
        self.num_consecutive_playing_steps = EnvironmentSteps(5000)
        self.act_for_full_episodes = True


class PPOAgentParameters(AgentParameters):
    def __init__(self):
        super().__init__(algorithm=PPOAlgorithmParameters(),
                         exploration={DiscreteActionSpace: CategoricalParameters(),
                                      BoxActionSpace: AdditiveNoiseParameters()},
                         memory=EpisodicExperienceReplayParameters(),
                         networks={"critic": PPOCriticNetworkParameters(), "actor": PPOActorNetworkParameters()})

    @property
    def path(self):
        return 'rl_coach.agents.ppo_agent:PPOAgent'


# Proximal Policy Optimization - https://arxiv.org/pdf/1707.06347.pdf
class PPOAgent(ActorCriticAgent):
    def __init__(self, agent_parameters, parent: Union['LevelManager', 'CompositeAgent']=None):
        super().__init__(agent_parameters, parent)

        # signals definition
        self.value_loss = self.register_signal('Value Loss')
        self.policy_loss = self.register_signal('Policy Loss')
        self.kl_divergence = self.register_signal('KL Divergence')
        self.total_kl_divergence_during_training_process = 0.0
        self.unclipped_grads = self.register_signal('Grads (unclipped)')

    @property
    def is_on_policy(self) -> bool:
        return True

    def fill_advantages(self, batch):
        batch = Batch(batch)
        network_keys = self.ap.network_wrappers['critic'].input_embedders_parameters.keys()

        # * Found not to have any impact *
        # current_states_with_timestep = self.concat_state_and_timestep(batch)

        current_state_values = self.networks['critic'].online_network.predict(batch.states(network_keys)).squeeze()
        total_returns = batch.n_step_discounted_rewards()
        # calculate advantages
        advantages = []
        if self.policy_gradient_rescaler == PolicyGradientRescaler.A_VALUE:
            advantages = total_returns - current_state_values
        elif self.policy_gradient_rescaler == PolicyGradientRescaler.GAE:
            # get bootstraps
            episode_start_idx = 0
            advantages = np.array([])
            # current_state_values[batch.game_overs()] = 0
            for idx, game_over in enumerate(batch.game_overs()):
                if game_over:
                    # get advantages for the rollout
                    value_bootstrapping = np.zeros((1,))
                    rollout_state_values = np.append(current_state_values[episode_start_idx:idx+1], value_bootstrapping)

                    rollout_advantages, _ = \
                        self.get_general_advantage_estimation_values(batch.rewards()[episode_start_idx:idx+1],
                                                                     rollout_state_values)
                    episode_start_idx = idx + 1
                    advantages = np.append(advantages, rollout_advantages)
        else:
            screen.warning("WARNING: The requested policy gradient rescaler is not available")

        # standardize
        advantages = (advantages - np.mean(advantages)) / np.std(advantages)

        # TODO: this will be problematic with a shared memory
        for transition, advantage in zip(self.memory.transitions, advantages):
            transition.info['advantage'] = advantage

        self.action_advantages.add_sample(advantages)

    def train_value_network(self, dataset, epochs):
        loss = []
        batch = Batch(dataset)
        network_keys = self.ap.network_wrappers['critic'].input_embedders_parameters.keys()

        # * Found not to have any impact *
        # add a timestep to the observation
        # current_states_with_timestep = self.concat_state_and_timestep(dataset)

        mix_fraction = self.ap.algorithm.value_targets_mix_fraction
        total_returns = batch.n_step_discounted_rewards(True)
        for j in range(epochs):
            curr_batch_size = batch.size
            if self.networks['critic'].online_network.optimizer_type != 'LBFGS':
                curr_batch_size = self.ap.network_wrappers['critic'].batch_size
            for i in range(batch.size // curr_batch_size):
                # split to batches for first order optimization techniques
                current_states_batch = {
                    k: v[i * curr_batch_size:(i + 1) * curr_batch_size]
                    for k, v in batch.states(network_keys).items()
                }
                total_return_batch = total_returns[i * curr_batch_size:(i + 1) * curr_batch_size]
                old_policy_values = force_list(self.networks['critic'].target_network.predict(
                    current_states_batch).squeeze())
                if self.networks['critic'].online_network.optimizer_type != 'LBFGS':
                    targets = total_return_batch
                else:
                    current_values = self.networks['critic'].online_network.predict(current_states_batch)
                    targets = current_values * (1 - mix_fraction) + total_return_batch * mix_fraction

                inputs = copy.copy(current_states_batch)
                for input_index, input in enumerate(old_policy_values):
                    name = 'output_0_{}'.format(input_index)
                    if name in self.networks['critic'].online_network.inputs:
                        inputs[name] = input

                value_loss = self.networks['critic'].online_network.accumulate_gradients(inputs, targets)

                self.networks['critic'].apply_gradients_to_online_network()
                if isinstance(self.ap.task_parameters, DistributedTaskParameters):
                    self.networks['critic'].apply_gradients_to_global_network()
                self.networks['critic'].online_network.reset_accumulated_gradients()

                loss.append([value_loss[0]])
        loss = np.mean(loss, 0)
        return loss

    def concat_state_and_timestep(self, dataset):
        current_states_with_timestep = [np.append(transition.state['observation'], transition.info['timestep'])
                                        for transition in dataset]
        current_states_with_timestep = np.expand_dims(current_states_with_timestep, -1)
        return current_states_with_timestep

    def train_policy_network(self, dataset, epochs):
        loss = []
        for j in range(epochs):
            loss = {
                'total_loss': [],
                'policy_losses': [],
                'unclipped_grads': [],
                'fetch_result': []
            }
            #shuffle(dataset)
            for i in range(len(dataset) // self.ap.network_wrappers['actor'].batch_size):
                batch = Batch(dataset[i * self.ap.network_wrappers['actor'].batch_size:
                                      (i + 1) * self.ap.network_wrappers['actor'].batch_size])

                network_keys = self.ap.network_wrappers['actor'].input_embedders_parameters.keys()

                advantages = batch.info('advantage')
                actions = batch.actions()
                if not isinstance(self.spaces.action, DiscreteActionSpace) and len(actions.shape) == 1:
                    actions = np.expand_dims(actions, -1)

                # get old policy probabilities and distribution
                old_policy = force_list(self.networks['actor'].target_network.predict(batch.states(network_keys)))

                # calculate gradients and apply on both the local policy network and on the global policy network
                fetches = [self.networks['actor'].online_network.output_heads[0].kl_divergence,
                           self.networks['actor'].online_network.output_heads[0].entropy]

                inputs = copy.copy(batch.states(network_keys))
                inputs['output_0_0'] = actions

                # old_policy_distribution needs to be represented as a list, because in the event of discrete controls,
                # it has just a mean. otherwise, it has both a mean and standard deviation
                for input_index, input in enumerate(old_policy):
                    inputs['output_0_{}'.format(input_index + 1)] = input

                total_loss, policy_losses, unclipped_grads, fetch_result =\
                    self.networks['actor'].online_network.accumulate_gradients(
                        inputs, [advantages], additional_fetches=fetches)

                self.networks['actor'].apply_gradients_to_online_network()
                if isinstance(self.ap.task_parameters, DistributedTaskParameters):
                    self.networks['actor'].apply_gradients_to_global_network()

                self.networks['actor'].online_network.reset_accumulated_gradients()

                loss['total_loss'].append(total_loss)
                loss['policy_losses'].append(policy_losses)
                loss['unclipped_grads'].append(unclipped_grads)
                loss['fetch_result'].append(fetch_result)

                self.unclipped_grads.add_sample(unclipped_grads)

            for key in loss.keys():
                loss[key] = np.mean(loss[key], 0)

            if self.ap.network_wrappers['critic'].learning_rate_decay_rate != 0:
                curr_learning_rate = self.networks['critic'].online_network.get_variable_value(self.ap.learning_rate)
                self.curr_learning_rate.add_sample(curr_learning_rate)
            else:
                curr_learning_rate = self.ap.network_wrappers['critic'].learning_rate

            # log training parameters
            screen.log_dict(
                OrderedDict([
                    ("Surrogate loss", loss['policy_losses'][0]),
                    ("KL divergence", loss['fetch_result'][0]),
                    ("Entropy", loss['fetch_result'][1]),
                    ("training epoch", j),
                    ("learning_rate", curr_learning_rate)
                ]),
                prefix="Policy training"
            )

        self.total_kl_divergence_during_training_process = loss['fetch_result'][0]
        self.entropy.add_sample(loss['fetch_result'][1])
        self.kl_divergence.add_sample(loss['fetch_result'][0])
        return loss['total_loss']

    def update_kl_coefficient(self):
        # John Schulman takes the mean kl divergence only over the last epoch which is strange but we will follow
        # his implementation for now because we know it works well
        screen.log_title("KL = {}".format(self.total_kl_divergence_during_training_process))

        # update kl coefficient
        kl_target = self.ap.algorithm.target_kl_divergence
        kl_coefficient = self.networks['actor'].online_network.get_variable_value(
            self.networks['actor'].online_network.output_heads[0].kl_coefficient)
        new_kl_coefficient = kl_coefficient
        if self.total_kl_divergence_during_training_process > 1.3 * kl_target:
            # kl too high => increase regularization
            new_kl_coefficient *= 1.5
        elif self.total_kl_divergence_during_training_process < 0.7 * kl_target:
            # kl too low => decrease regularization
            new_kl_coefficient /= 1.5

        # update the kl coefficient variable
        if kl_coefficient != new_kl_coefficient:
            self.networks['actor'].online_network.set_variable_value(
                self.networks['actor'].online_network.output_heads[0].assign_kl_coefficient,
                new_kl_coefficient,
                self.networks['actor'].online_network.output_heads[0].kl_coefficient_ph)

        screen.log_title("KL penalty coefficient change = {} -> {}".format(kl_coefficient, new_kl_coefficient))

    def post_training_commands(self):
        if self.ap.algorithm.use_kl_regularization:
            self.update_kl_coefficient()

        # clean memory
        self.call_memory('clean')

    def train(self):
        loss = 0
        if self._should_train():
            for network in self.networks.values():
                network.set_is_training(True)

            for training_step in range(self.ap.algorithm.num_consecutive_training_steps):
                self.networks['actor'].sync()
                self.networks['critic'].sync()

                dataset = self.memory.transitions

                self.fill_advantages(dataset)

                # take only the requested number of steps
                dataset = dataset[:self.ap.algorithm.num_consecutive_playing_steps.num_steps]

                value_loss = self.train_value_network(dataset, 1)
                policy_loss = self.train_policy_network(dataset, 10)

                self.value_loss.add_sample(value_loss)
                self.policy_loss.add_sample(policy_loss)

            for network in self.networks.values():
                network.set_is_training(False)

            self.post_training_commands()
            self.training_iteration += 1
            self.update_log()  # should be done in order to update the data that has been accumulated * while not playing *
            return np.append(value_loss, policy_loss)

    def get_prediction(self, states):
        tf_input_state = self.prepare_batch_for_inference(states, "actor")
        return self.networks['actor'].online_network.predict(tf_input_state)