1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203
| import random import torch from torch import nn from torch import optim import torch.nn.functional as F
BATCH_SIZE = 32 CAPACITY = 10000
class Brain: def __init__(self, num_states, num_actions): self.num_actions = num_actions
self.memory = ReplayMemory(CAPACITY)
n_in, n_mid, n_out = num_states, 32, num_actions self.main_q_network = Net(n_in, n_mid, n_out) self.target_q_network = Net(n_in, n_mid, n_out) print(self.main_q_network)
self.optimizer = optim.Adam( self.main_q_network.parameters(), lr=0.0001)
self.td_error_memory = TDerrorMemory(CAPACITY)
def replay(self, episode): '''Experience Replayでネットワークの結合パラメータを学習'''
if len(self.memory) < BATCH_SIZE: return
self.batch, self.state_batch, self.action_batch, self.reward_batch, self.non_final_next_states = self.make_minibatch(episode)
self.expected_state_action_values = self.get_expected_state_action_values()
self.update_main_q_network()
def decide_action(self, state, episode): '''現在の状態に応じて、行動を決定する''' epsilon = 0.5 * (1 / (episode + 1))
if epsilon <= np.random.uniform(0, 1): self.main_q_network.eval() with torch.no_grad(): action = self.main_q_network(state).max(1)[1].view(1, 1)
else: action = torch.LongTensor( [[random.randrange(self.num_actions)]])
return action
def make_minibatch(self, episode): '''2. ミニバッチの作成'''
if episode < 30: transitions = self.memory.sample(BATCH_SIZE) else: indexes = self.td_error_memory.get_prioritized_indexes(BATCH_SIZE) transitions = [self.memory.memory[n] for n in indexes]
batch = Transition(*zip(*transitions))
state_batch = torch.cat(batch.state) action_batch = torch.cat(batch.action) reward_batch = torch.cat(batch.reward) non_final_next_states = torch.cat([s for s in batch.next_state if s is not None])
return batch, state_batch, action_batch, reward_batch, non_final_next_states
def get_expected_state_action_values(self): '''3. 教師信号となるQ(s_t, a_t)値を求める'''
self.main_q_network.eval() self.target_q_network.eval()
self.state_action_values = self.main_q_network( self.state_batch).gather(1, self.action_batch)
non_final_mask = torch.ByteTensor(tuple(map(lambda s: s is not None, self.batch.next_state))) next_state_values = torch.zeros(BATCH_SIZE) a_m = torch.zeros(BATCH_SIZE).type(torch.LongTensor)
a_m[non_final_mask] = self.main_q_network(self.non_final_next_states).detach().max(1)[1]
a_m_non_final_next_states = a_m[non_final_mask].view(-1, 1)
next_state_values[non_final_mask] = self.target_q_network( self.non_final_next_states).gather(1, a_m_non_final_next_states).detach().squeeze()
expected_state_action_values = self.reward_batch + GAMMA * next_state_values
return expected_state_action_values
def update_main_q_network(self): '''4. 結合パラメータの更新'''
self.main_q_network.train()
loss = F.smooth_l1_loss(self.state_action_values, self.expected_state_action_values.unsqueeze(1))
self.optimizer.zero_grad() loss.backward() self.optimizer.step()
def update_target_q_network(self): '''Target Q-NetworkをMainと同じにする''' self.target_q_network.load_state_dict(self.main_q_network.state_dict())
def update_td_error_memory(self): '''TD誤差メモリに格納されているTD誤差を更新する'''
self.main_q_network.eval() self.target_q_network.eval()
transitions = self.memory.memory batch = Transition(*zip(*transitions))
state_batch = torch.cat(batch.state) action_batch = torch.cat(batch.action) reward_batch = torch.cat(batch.reward) non_final_next_states = torch.cat([s for s in batch.next_state if s is not None])
state_action_values = self.main_q_network(state_batch).gather(1, action_batch)
non_final_mask = torch.ByteTensor(tuple(map(lambda s: s is not None, batch.next_state)))
next_state_values = torch.zeros(len(self.memory)) a_m = torch.zeros(len(self.memory)).type(torch.LongTensor)
a_m[non_final_mask] = self.main_q_network(non_final_next_states).detach().max(1)[1]
a_m_non_final_next_states = a_m[non_final_mask].view(-1, 1)
next_state_values[non_final_mask] = self.target_q_network( non_final_next_states).gather(1, a_m_non_final_next_states).detach().squeeze()
td_errors = (reward_batch + GAMMA * next_state_values) - state_action_values.squeeze()
self.td_error_memory.memory = td_errors.detach().numpy().tolist()
|