游戏网站模板,网站开发培训程序员,国外免费推广网站,学做网站教程人工智能例子汇总#xff1a;AI常见的算法和例子-CSDN博客 特性GRULSTM计算效率更快#xff0c;参数更少相对较慢#xff0c;参数更多结构复杂度只有两个门#xff08;更新门和重置门#xff09;三个门#xff08;输入门、遗忘门、输出门#xff09;处理长时依赖一般适… 人工智能例子汇总AI常见的算法和例子-CSDN博客 特性GRULSTM计算效率更快参数更少相对较慢参数更多结构复杂度只有两个门更新门和重置门三个门输入门、遗忘门、输出门处理长时依赖一般适用于中等长度依赖更适合处理超长时序依赖训练速度训练更快梯度更稳定训练较慢占用更多内存
例子
import torch
import torch.nn as nn
import torch.optim as optim
import random
import matplotlib.pyplot as plt# 迷宫环境5×5
class MazeEnv:def __init__(self, size5):self.size sizeself.state (0, 0) # 起点self.goal (size-1, size-1) # 终点self.actions [(0,1), (0,-1), (1,0), (-1,0)] # 右、左、下、上def reset(self):self.state (0, 0) # 重置起点return self.statedef step(self, action):dx, dy self.actions[action]x, y self.statenx, ny max(0, min(self.size-1, xdx)), max(0, min(self.size-1, ydy))reward 1 if (nx, ny) self.goal else -0.1done (nx, ny) self.goalself.state (nx, ny)return (nx, ny), reward, done# GRU 策略网络
class GRUPolicy(nn.Module):def __init__(self, input_size, hidden_size, output_size):super(GRUPolicy, self).__init__()self.gru nn.GRU(input_size, hidden_size, batch_firstTrue)self.fc nn.Linear(hidden_size, output_size)def forward(self, x, hidden):out, hidden self.gru(x, hidden)out self.fc(out[:, -1, :]) # 只取最后时间步return out, hidden# 训练参数
env MazeEnv(size5)
policy GRUPolicy(input_size2, hidden_size16, output_size4)
optimizer optim.Adam(policy.parameters(), lr0.01)
loss_fn nn.CrossEntropyLoss()# 训练
num_episodes 500
epsilon 1.0 # 初始的ε值控制探索的概率
epsilon_min 0.01 # 最小ε值
epsilon_decay 0.995 # ε衰减率
best_path [] # 用于存储最佳路径for episode in range(num_episodes):state env.reset()hidden torch.zeros(1, 1, 16) # GRU 初始状态states, actions, rewards [], [], []logits_list [] for _ in range(20): # 最多 20 步state_tensor torch.tensor([[state[0], state[1]]], dtypetorch.float32).unsqueeze(0)logits, hidden policy(state_tensor, hidden)logits_list.append(logits)# ε-greedy 策略if random.random() epsilon:action random.choice(range(4)) # 随机选择动作else:action torch.argmax(logits, dim1).item() # 选择最大值对应的动作next_state, reward, done env.step(action)states.append(state)actions.append(action)rewards.append(reward)if done:print(fEpisode {episode} - Reached Goal!)# 找到最优路径best_path states [next_state] # 当前 episode 的路径breakstate next_state# 计算损失logits torch.cat(logits_list, dim0) # (T, 4)action_tensor torch.tensor(actions, dtypetorch.long) # (T,)loss loss_fn(logits, action_tensor) optimizer.zero_grad()loss.backward()optimizer.step()# 衰减 εepsilon max(epsilon_min, epsilon * epsilon_decay)if episode % 100 0:print(fEpisode {episode}, Loss: {loss.item():.4f}, Epsilon: {epsilon:.4f})# 确保 best_path 已经记录
if len(best_path) 0:print(No path found during training.)
else:print(fBest path: {best_path})# 测试路径只绘制最佳路径
fig, ax plt.subplots(figsize(6,6))# 初始化迷宫图
maze [[0 for _ in range(5)] for _ in range(5)] # 5×5 迷宫
ax.imshow(maze, cmapcoolwarm, originupper)# 画网格
ax.set_xticks(range(5))
ax.set_yticks(range(5))
ax.grid(True, colorblack, linewidth0.5)# 画出最佳路径红色
for (x, y) in best_path:ax.add_patch(plt.Rectangle((y, x), 1, 1, colorred, alpha0.8))# 画起点和终点
ax.text(0, 0, S, hacenter, vacenter, fontsize14, colorwhite, fontweightbold)
ax.text(4, 4, G, hacenter, vacenter, fontsize14, colorwhite, fontweightbold)plt.title(GRU RL Agent - Best Path)
plt.show()
