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 g' l6 l8 B! ]' w1 {+ m/ L
import numpy as np# r6 l/ P. F0 D" ^( l/ M
4 j' }6 F) m( G7 P
# 初始化环境
9 G# }% N1 R6 K/ s
env = gym.make('CartPole-v1')
4 W: Q* p. O0 H0 p) P% [+ D
n_actions = env.action_space.n. [7 B0 s7 e- _ V+ a. _
n_states = env.observation_space.shape[0]! u! i$ \& ^ Z. B
" n, s6 Y! u V
# 初始化Q表
7 T e/ K/ r+ M2 P/ m9 y1 W2 Y
Q = np.zeros((n_states, n_actions))' ~0 t4 c4 x0 T# t& R
: r& Y2 ?8 |* g% a% Q8 L2 ~7 V
# 超参数
8 J; X* ^3 W1 r' l
alpha = 0.1 # 学习率
/ { o, S( X$ D1 M4 d3 D
gamma = 0.99 # 折扣因子0 r7 v2 F- P1 C% t) H: X
epsilon = 0.1 # 探索率
+ x0 Q) D2 a' n6 t0 k* C; Z2 h
! X4 A" n2 j0 c
# 训练过程
! Q5 ?, `4 `( U8 R% x$ {) v6 J' g
for episode in range(1000):
1 J7 L3 C! m( I. |2 Z2 u- Z ^: l" p
state = env.reset()0 p& o/ m- Y$ E
done = False
( O, L8 n. q+ e, }
U* B" q& Q6 z7 a1 H( S
while not done:
" E ~" D! Q, m
# epsilon-贪婪策略进行动作选择
+ J" t; X) l6 z6 [5 y
if np.random.rand() < epsilon:
2 h9 n T5 p8 g# c
action = env.action_space.sample() # 探索. ~; a2 f& {+ H' c' j( D# C6 C9 L
else:
; Z7 Y, ~. Q- A& Q
action = np.argmax(Q[state, :]) # 利用# O, L/ E) `$ B8 G* D X+ [
: h$ c- N* T! D5 |0 Z
# 执行动作
k$ A' Z! _9 @# r; T
next_state, reward, done, _ = env.step(action). f, ]: j" D5 J0 i6 M6 ~
) f+ P+ V1 l+ b5 A/ v7 d8 R+ ^1 C
# Q表更新( J2 ]7 g; }7 u$ a, U! `7 r0 k
Q[state, action] = Q[state, action] + alpha * (reward + gamma * np.max(Q[next_state, :]) - Q[state, action])
" q+ O) `% D; x7 h- U. v3 S
- Y6 A9 _* B$ B9 @+ S2 E9 b
state = next_state
; l7 j8 e, s. B! A: p& {
1 N" j9 X1 h: z U& L! h6 T
# 测试智能体

复制代码