politer-selfplay.py

# # source: https://web.stanford.edu/~surag/posts/alphazero.html

# def policyIterSP(game):
#     nnet = initNNet()                                       # initialise random neural network
#     examples = []
#     for i in range(numIters):
#         for e in range(numEps):
#             examples += executeEpisode(game, nnet)          # collect examples from this game
#         new_nnet = trainNNet(examples)
#         frac_win = pit(new_nnet, nnet)                      # compare new net with previous net
#         if frac_win > threshold:
#             nnet = new_nnet                                 # replace with new net
#     return nnet

# def executeEpisode(game, nnet):
#     examples = []
#     s = game.startState()
#     mcts = MCTS()                                           # initialise search tree

#     while True:
#         for _ in range(numMCTSSims):
#             mcts.search(s, game, nnet)
#         examples.append([s, mcts.pi(s), None])              # rewards can not be determined yet
#         a = random.choice(len(mcts.pi(s)), p=mcts.pi(s))    # sample action from improved policy
#         s = game.nextState(s,a)
#         if game.gameEnded(s):
#             examples = assignRewards(examples, game.gameReward(s))
#             return examples