Multiagent cooperation and competition with deep reinforcement learning

论文复现 :

tensorflow_2player_pong

论文详述

Multiagent cooperation and competition with deep reinforcement learning

pong game-two agents

• 基础模型：pong game, two agents
• 算法结构：dqn
  • reward：scoring:(-1,1) conceding(-1)
    未击中球得-1，击中球得分between (-1,1)
    双方均击中球得分0，游戏继续

reward

• 训练参数
  • 50 epochs, 250000 time steps each.
  • exploration rate: 1.0 to 0.05(in the 1000000 time steps) and stays fixed at that value

parameters.png

• 结果分析

  • 是否收敛:monitor average maximal Q-values of 500 randomly selected game situations, set aside before training begins

    
    
    Q values

  • 训练效果反馈:

    • Average paddle-bounces per point 在一方得分前球在players间来回的次数
    • Average wall-bounces per paddle-bounce 球在到达一方前撞墙的次数
    • Average serving time per point 球丢了以后players restart game的反应时间(一些rewarding scheme下players不希望重启游戏，serving time很长，如p = -1)

结果分析

• scoring = -1时，双方为合作状态（均不希望球掉落）
  最终双方均升至页面最上方，球水平传来传去
  合作模式video-youtube
  
  1.png
• scoring = 1时，双方为竞争模式(希望自己多得分)
  竞争模式video-youtube
  
  2.png
• p range from -1 to 1

3.png

• multiplayer dqn vs single-player
  (score表示a胜b的得分)

  
  
  4

---

本文遵守知识共享协议：署名-非商业性使用-相同方式共享 (BY-NC-SA)及协议
转载请注明：作者空空格格，首发 Jianshu.com