经典控制任务（Q学习/策略梯度） - 数学建模社区-数学中国

import gym
9 a. E! \- o3 v# N. K
import numpy as np- X# n* p" s' b g& e" U
. a& U+ _9 o$ [8 x) V) {
# 初始化环境
7 S! t t0 d) k3 L k0 w
env = gym.make('CartPole-v1')
0 w1 w; _, P- q
n_actions = env.action_space.n; x8 A* t3 t8 v, ^* I
n_states = env.observation_space.shape[0]
1 R. d6 d! e: s9 s5 A5 J; r9 l
6 y1 B0 N% r4 U+ ?8 A5 K4 q
# 初始化Q表4 e }: o* b5 g- I( h2 ~" P2 w$ n/ z
Q = np.zeros((n_states, n_actions))* ^1 H) |* f1 S# i% b: X1 Z( q5 S
9 l1 Z( a& E/ ~* ?6 F
# 超参数
0 U: E) w) K" z; {' `! w
alpha = 0.1 # 学习率
gamma = 0.99 # 折扣因子% V6 j' H3 g; G2 W6 [
epsilon = 0.1 # 探索率
2 v6 e3 Y: S! E
" T5 i3 @+ C3 i6 p7 Q' O; B. R
# 训练过程# B3 A$ ]1 v0 D& i9 _4 [/ h0 }
for episode in range(1000):
$ e: T) A7 H4 _3 D! e8 q
state = env.reset(); Q. N, `! Z. Y1 O+ I
done = False
$ e$ @+ a5 Y; r: r& |
. _/ O+ f5 ?, q: L& u3 }
while not done:
+ M% o4 B* U4 v
# epsilon-贪婪策略进行动作选择* p$ W3 q6 {5 ^% [9 |9 U% w
if np.random.rand() < epsilon:6 Z9 a# G$ l# T$ C F3 P M6 F) x2 s
action = env.action_space.sample() # 探索
2 d0 W4 m! Z: m0 R0 L
else:1 G5 p2 B J O; h4 s# s2 |! v7 H
action = np.argmax(Q[state, :]) # 利用
9 F* R" N" }: h5 h: W7 j
0 k- V" A& H# F" a. e# J \
# 执行动作, h ?$ N. w3 h) {
next_state, reward, done, _ = env.step(action)! r) e" D% Z* ~7 z. h1 R( y
7 h% }+ M, f6 P' R1 Q( @
# Q表更新' `1 ?1 }# A5 S& `$ X2 d
Q[state, action] = Q[state, action] + alpha * (reward + gamma * np.max(Q[next_state, :]) - Q[state, action])
' {8 e+ Z* g4 W; U3 Z. C/ B% ~
- d4 j2 f& Y! N/ x* L+ g: W
state = next_state
1 V7 I" }% C: L+ q, o' q) [1 }
$ w, b `3 j s9 \. e0 K4 @
# 测试智能体

复制代码