Aim: Train a model to properly play vintage video games...
Deep Q-learning Algo~
Very short Brief of Notations:
{A,pi(Policy),Q(quality of action-at a state),R ((s,a,s') - Reward, s state doing a to go to s' and get a specific r)}
So, if we want to train a model to play a video game like master. Modules are to be implemented as minimum, listed. below:
- a class that can catch enough frames(typically consequtive) for game env analysis -> might need preprocessing to lower the memory overhead
- a class of NN based model for training, weights init/update/storage/write/fork/reset; also the actions in a single play is recorded for optimization
- a class that utilize the possible actions and abstrct to humble level to do anything player is going to do w/o generative issue at the beginning(can go general when model matured)
- game to model/pre-processing module
This is just the minimum...
标签:play,implementation,actions,game,learning,RL,model,class From: https://www.cnblogs.com/selfmade-Henderson/p/17672711.html