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
- V7 u, P) u) F
import numpy as np
- W! ^; p9 o" Y [: J$ [
1 L. ]8 L9 X1 f _# J4 M
# 初始化环境
& D/ R* }7 R9 |: Q8 i: Z- ^
env = gym.make('CartPole-v1')$ I7 C2 v+ ?2 w" ^
n_actions = env.action_space.n
6 a8 d2 W( C8 w2 F P4 p
n_states = env.observation_space.shape[0]
+ ^6 n5 V2 U! a6 S) h) ] J
- I6 o( i7 h9 {' l5 ~( a. B
# 初始化Q表1 ~3 B, K% U. u# B7 o3 O" T" ?
Q = np.zeros((n_states, n_actions))4 o. ~1 s! P! }( D" P
/ ]9 a# i. ]( W% a* A- E3 Y
# 超参数
% c/ r9 e5 l# h' {* ~* e
alpha = 0.1 # 学习率: L4 C. c4 M/ y9 @% [6 }5 N
gamma = 0.99 # 折扣因子
; R6 a5 E2 x3 X) @
epsilon = 0.1 # 探索率; ?" `, c2 ~$ f y4 W
: ^& z- e* v# M
# 训练过程0 L: `/ H' ]+ `/ V& p, o' p
for episode in range(1000):9 r: ^- K5 }7 J
state = env.reset()0 d* Y# I- _& Y; _, Z
done = False' _" Z4 ]; Q: S1 J
. n- J, ~, T8 n$ K( j K. G
while not done:
& n$ ?7 r5 u9 p r& `7 H6 ]- r
# epsilon-贪婪策略进行动作选择
# e3 X+ H5 q: m! C/ [* X
if np.random.rand() < epsilon:
! J% }) A Q8 `, F1 p& w
action = env.action_space.sample() # 探索' I6 M7 u# Z5 m7 |0 t
else:! r4 g9 Z2 S+ r/ E- S
action = np.argmax(Q[state, :]) # 利用4 ^4 y) C# x6 E8 d* `
0 p7 N: m! l# ^+ y. u; |) [
# 执行动作
1 U! i: M' b" R6 H+ c) R
next_state, reward, done, _ = env.step(action)/ Y O6 v: j. g7 v8 K `3 p: v t
* D# R& o3 ?8 t2 \* a1 I! v
# Q表更新- Y2 ], j4 W _5 y/ ~, c' |
Q[state, action] = Q[state, action] + alpha * (reward + gamma * np.max(Q[next_state, :]) - Q[state, action])
+ W0 [, c% i- z2 P1 e
5 y! S1 v' B& S; [$ E: ?
state = next_state
2 C& `& f* j: G& L: }, }
2 R- `3 ]; k+ g
# 测试智能体

复制代码