suragnair

def search(s, game, nnet):
    if game.gameEnded(s): return -game.gameReward(s)

    if s not in visited:
        visited.add(s)
        P[s], v = nnet.predict(s)
        return -v
  
    max_u, best_a = -float("inf"), -1
    for a in game.getValidActions(s):

suragnair

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