跳到主要內容

增強式學習

 

 迴力球遊戲－ATARI

 



 

賽車遊戲DQN－ATARI


賽車遊戲－TORCS








Ref:
    李宏毅老師 YOUTUBE DRL 1-3

On-policy VS Off-policy

On-policy
    The agent learned and the agent interacting with the environment is the same
    阿光自已下棋學習
Off-policy
    The agent learned and the agent interacting with the environment is different

    佐助下棋，阿光在旁邊看



Add a baseline:
    It is possible that R is always positive
    So R subtract a expectation value


Policy in "Policy Gradient" means output action, like left/right/fire


gamma-discounted rewards:
時間愈遠的貢獻，降低其權重

Reward Function & Action is defined in prior to training

MC v.s. TD
MC 蒙弟卡羅: critic after episode end : larger variance(cuz conditions differ a lot in every episode), unbiased (judge until episode end, more fair)
TD: Temporal-difference approach: critic during episode :smaller variance, biased maybe



atari : a3c  => gym
torcs : ddpg => gym-torcs


PPO
   easy code 
   easy tune
   sample efficient


Replay Buffer:
Put the experience into buffer : St, at, rt, St+1
cuz interaction with environment cost more time than training
玩遊戲的時間通常比TRAIN NETWORK花時間

留言

張貼留言

這個網誌中的熱門文章

A3C in ATARI Pong-V0

ATARI PONG 對戰模式，左邊為遊戲程式，右邊為訓練中的A3C模型。一局以21分決勝負，對手MISS 一球得一分。從LOG可以看出，A3C模型從最初全敗的輸21分，經過2小時左右的TRAINING，已經逆轉至幾乎每局都勝利，偶爾甚至勝出高達13分。 底下為TRAINING A3C MODEL過程的LOG， (base) frank@viper1:~/a3c$ python main.py --env-name "Pong-v0" --num-processes 8 Time 00h 00m 10s, episode reward -21.0, episode length 1026 Time 00h 01m 18s, episode reward -21.0, episode length 1020 Time 00h 02m 26s, episode reward -21.0, episode length 1029 Time 00h 03m 35s, episode reward -21.0, episode length 1023 Time 00h 04m 42s, episode reward -21.0, episode length 1014 Time 00h 05m 50s, episode reward -21.0, episode length 1087 Time 00h 07m 00s, episode reward -21.0, episode length 1359 Time 00h 08m 14s, episode reward -16.0, episode length 1922 Time 00h 09m 30s, episode reward -14.0, episode length 2220 Time 00h 10m 48s, episode reward -15.0, episode length 2431 Time 00h 12m 04s, episode reward -15.0, episode length 2211 Time 00h 13m 23s, episode reward -8.0, episode length 2600 Time 00h 14m 38s, episod...
張貼留言
閱讀完整內容

DDPG in Torcs within Docker Container

Dockerfile: FROM tensorflow/tensorflow:0.10.0-gpu WORKDIR /home/frank/old_gym_torcs ADD . /home/old_gym_torcs  RUN apt update RUN apt install -y vim xautomation torcs RUN apt-get install -y libjpeg-dev cmake swig python-pyglet python3-opengl libboost-all-dev \         libsdl2-2.0.0 libsdl2-dev libglu1-mesa libglu1-mesa-dev libgles2-mesa-dev \         freeglut3 xvfb libav-tools RUN pip install gym RUN pip install keras==1.1.0 ENV PATH="/usr/games:${PATH}" CMD ["/bin/bash"] viper1 $ docker run --runtime=nvidia -it -e DISPLAY=$DISPLAY -v /tmp/.X11-unix:/tmp/.X11-unix -v /home/frank/old_gym_torcs:/home/old_gym_torcs -v /var/run/docker.sock:/var/run/docker.sock -v /home/frank/gym_torcs:/home/gym_torcs -v /home/frank/gym:/home/gym -p 3101:3101 --workdir /home/old_gym_torcs -p 8888:8888 ddpgfrk:tf0.10 /bin/bash (grey part is not necessary) After  $ docker commit id ddpg:tf0.10.0-gpu viper1 $ docker r...
張貼留言
閱讀完整內容

Loss function

why cross entroy is more suitable for categorical classcification? ref: https://www.youtube.com/watch?v=Li5sVEXTIJw
張貼留言
閱讀完整內容