analyzer_network_training.py

import numpy as np
import random
from holdem import Analyzer, NeuralNetwork, HoldemAI
from deuces.deuces import Card, Deck, Evaluator


if __name__ = '__main__':
    a = Analyzer()
    a.monte_carlo_rounds = 2000

    nn = NeuralNetwork([206,206,100,1], 'analyzer_network', 0.1)

    deck = Deck()
    hand = []
    community = []
    opponents = 0
    game_stage = [0,3,4,5]

    for i in range(1000000):
        community = []
        deck.shuffle()
        a.reset()

        hand = deck.draw(2)
        for _ in range(game_stage[random.randint(0,3)]):
            community.append(deck.draw(1))
        opponents = random.randint(1,7)

        a.set_num_opponents(opponents)
        a.set_pocket_cards(*hand)
        for card in community:
            a.community_card(card)

        mc_response = a.analyze()
        # print('mc response', mc_response)

        hand_bin = [j for i in [HoldemAI.card_to_binlist(c) for c in hand] for j in i]
        community = community + [0]*(5-len(community))
        comm_bin = [j for i in [HoldemAI.card_to_binlist(c) for c in community] for j in i]
        opponents_bin = HoldemAI.bin_to_binlist(bin(opponents)[2:].zfill(3))

        inputs = HoldemAI.center_bin(hand_bin + comm_bin + opponents_bin)

        # print('nn response', (1+nn.activate(inputs))/2)
        print('error: ', 100*((nn.activate(inputs)+1)/2 - mc_response))
        nn.update_weights_linesearch(inputs, 2*mc_response-1)

        if (i % 100) == 0:
            print('SAVING ON ITERATION: ', i)
            nn.save()