rltorch/rltorch/agents/QEPAgent.py

121 lines
5.9 KiB
Python

from copy import deepcopy
import collections
import numpy as np
import torch
from torch.distributions import Categorical
import rltorch
import rltorch.memory as M
# Q-Evolutionary Policy Agent
# Maximizes the policy with respect to the Q-Value function.
# Since function is non-differentiabile, depends on the Evolutionary Strategy algorithm
class QEPAgent:
def __init__(self, policy_net, value_net, memory, config, target_value_net = None, logger = None):
self.policy_net = policy_net
assert isinstance(self.policy_net, rltorch.network.ESNetwork)
self.policy_net.fitness = self.fitness
self.value_net = value_net
self.target_value_net = target_value_net
self.memory = memory
self.config = deepcopy(config)
self.logger = logger
self.policy_skip = 4
def fitness(self, policy_net, value_net, state_batch):
batch_size = len(state_batch)
action_probabilities = policy_net(state_batch)
action_size = action_probabilities.shape[1]
distributions = list(map(Categorical, action_probabilities))
actions = torch.stack([d.sample() for d in distributions])
with torch.no_grad():
state_values = value_net(state_batch)
obtained_values = state_values.gather(1, actions.view(len(state_batch), 1)).squeeze(1)
# return -obtained_values.mean().item()
entropy_importance = 0.01 # Entropy accounting for 1% of loss seems to work well
value_importance = 1 - entropy_importance
# entropy_loss = (action_probabilities * torch.log2(action_probabilities)).sum(1) # Standard entropy loss from information theory
entropy_loss = (action_probabilities - torch.tensor(1 / action_size).repeat(len(state_batch), action_size)).abs().sum(1)
return (entropy_importance * entropy_loss - value_importance * obtained_values).mean().item()
def learn(self, logger = None):
if len(self.memory) < self.config['batch_size']:
return
if (isinstance(self.memory, M.PrioritizedReplayMemory)):
weight_importance = self.config['prioritized_replay_weight_importance']
# If it's a scheduler then get the next value by calling next, otherwise just use it's value
beta = next(weight_importance) if isinstance(weight_importance, collections.Iterable) else weight_importance
minibatch = self.memory.sample(self.config['batch_size'], beta = beta)
state_batch, action_batch, reward_batch, next_state_batch, not_done_batch, importance_weights, batch_indexes = M.zip_batch(minibatch, priority = True)
else:
minibatch = self.memory.sample(self.config['batch_size'])
state_batch, action_batch, reward_batch, next_state_batch, not_done_batch = M.zip_batch(minibatch)
# Send to their appropriate devices
state_batch = state_batch.to(self.value_net.device)
action_batch = action_batch.to(self.value_net.device)
reward_batch = reward_batch.to(self.value_net.device).float()
next_state_batch = next_state_batch.to(self.value_net.device)
not_done_batch = not_done_batch.to(self.value_net.device)
state_values = self.value_net(state_batch)
obtained_values = state_values.gather(1, action_batch.view(self.config['batch_size'], 1))
with torch.no_grad():
# Use the target net to produce action values for the next state
# and the regular net to select the action
# That way we decouple the value and action selecting processes (DOUBLE DQN)
not_done_size = not_done_batch.sum()
next_state_values = torch.zeros_like(state_values, device = self.value_net.device)
if self.target_value_net is not None:
next_state_values[not_done_batch] = self.target_value_net(next_state_batch[not_done_batch])
next_best_action = self.value_net(next_state_batch[not_done_batch]).argmax(1)
else:
next_state_values[not_done_batch] = self.value_net(next_state_batch[not_done_batch])
next_best_action = next_state_values[not_done_batch].argmax(1)
best_next_state_value = torch.zeros(self.config['batch_size'], device = self.value_net.device)
best_next_state_value[not_done_batch] = next_state_values[not_done_batch].gather(1, next_best_action.view((not_done_size, 1))).squeeze(1)
expected_values = (reward_batch + (self.config['discount_rate'] * best_next_state_value)).unsqueeze(1)
if (isinstance(self.memory, M.PrioritizedReplayMemory)):
value_loss = (torch.as_tensor(importance_weights, device = self.value_net.device) * ((obtained_values - expected_values)**2).squeeze(1)).mean()
else:
value_loss = F.mse_loss(obtained_values, expected_values)
if self.logger is not None:
self.logger.append("Loss/Value", value_loss.item())
self.value_net.zero_grad()
value_loss.backward()
self.value_net.clamp_gradients()
self.value_net.step()
if self.target_value_net is not None:
if 'target_sync_tau' in self.config:
self.target_value_net.partial_sync(self.config['target_sync_tau'])
else:
self.target_value_net.sync()
if (isinstance(self.memory, M.PrioritizedReplayMemory)):
td_error = (obtained_values - expected_values).detach().abs()
self.memory.update_priorities(batch_indexes, td_error)
## Policy Training
if self.policy_skip > 0:
self.policy_skip -= 1
return
self.policy_skip = 4
if self.target_value_net is not None:
self.policy_net.calc_gradients(self.target_value_net, state_batch)
else:
self.policy_net.calc_gradients(self.value_net, state_batch)
# self.policy_net.clamp_gradients()
self.policy_net.step()