tic_tac_toe/tic_tac_toe.py

import numpy as np
from numpy.random import uniform

float_formatter = "{:.3f}".format
np.set_printoptions(formatter={'float_kind':float_formatter})


def softmax(values: np.array) -> np.array:
    result = values / sum(values)
    result = np.sort(result)
    return result


def sample(values: np.array, with_debug=False):
    # Normalize and sort
    probs = softmax(values + 1.0E-6)
    z = uniform()
    if with_debug:
        print(f"Probs={probs}")
        print(f"Z={z:.3f}")
    index = None
    p_sum = 0
    for idx, p in enumerate(probs):
        p_sum += p
        if z <= p_sum:
            index = idx
            break

    probs_max = np.max(probs)
    is_exploration = probs[index] < probs_max
    return index, is_exploration


def test_sample():
    p = np.array([0.1, 0.1, 0.3, 0.5])
    c = np.array([0, 0, 0, 0])

    for i in range(0, 1000):
        index = sample(p)
        c[index] += 1

    print(c)


def to_state_string(state, state_nex=None):
    sp = '      '
    sp_arrow = '  =>  '
    sp_ = [sp, sp_arrow, sp]

    def col(str_in, state):
        str_out = str_in
        for c in range(0, 3):
            char = state[3*r+c]
            if char == '-':
                char = ' '
            str_out += "|" + char
        str_out += '|'
        return str_out

    state_str = ''
    for r in range(0, 3):
        state_str = col(state_str, state)

        if state_nex is not None:
            state_str += sp_[r]
            state_str = col(state_str, state_nex)

        if r != 2:
            state_str += '\x0A'

    return state_str


class APlayer(object):
    def __init__(self, mark):
        self.mark = mark
        self.state = None
        self.state_last = None

    def move(self, state):
        return state, False

    def has_won(self, state):
        ref = self.mark + self.mark + self.mark

        substr = state[0:3]
        if substr in ref:
            return True

        substr = state[3:6]
        if substr in ref:
            return True

        substr = state[6:9]
        if substr in ref:
            return True

        substr = state[0:9:3]
        if substr in ref:
            return True

        substr = state[1:9:3]
        if substr in ref:
            return True

        substr = state[2:9:3]
        if substr in ref:
            return True

        substr = state[0:9:4]
        if substr in ref:
            return True

        substr = state[6:0:-2]
        if substr in ref:
            return True

        return False


class MachinePlayer(APlayer):
    def __init__(self, mark='X', with_debug=False):
        APlayer.__init__(self, mark)
        self.with_debug = with_debug
        self.values = {}

    def get_value(self, state):
        try:
            result = self.values[state]
        except KeyError:
            result = 0

        return result

    def set_value(self, value):
        if self.state_last is not None:
            self.values[self.state_last] = value

    @staticmethod
    def get_potential_moves(state) -> np.array:
        indices = [idx for idx, s in enumerate(state) if '-' in s]
        return np.array(indices)

    def move(self, state):
        values = np.array([])
        # get possible move
        moves = self.get_potential_moves(state)
        can_move = moves.size > 0

        if can_move:
            for field in moves:
                # crate hypothetical next state
                state_next = self.state_from_move(state, field)
                # evaluate value
                value = self.get_value(state_next)
                values = np.append(values, value)

            index, is_exp = sample(values, self.with_debug)
            field = moves[index]
            if self.with_debug:
                print(f"{self.mark}: Chose {index}, is_exp={is_exp}")
            self.state_last = self.state
            self.state = self.state_from_move(state, field)

            # Learn
            if not is_exp and self.state_last is not None:
                v0 = self.get_value(self.state_last)
                v1 = self.get_value(self.state)
                d = max(0, v1-v0)
                if d > 0:
                    self.set_value(0.1*d)
                    if self.with_debug:
                        print(f"{self.mark}: Learned {d:0.3f}")

        return self.state, can_move

    def state_from_move(self, state, field):
        return state[:field] + self.mark + state[field + 1:]


def play(players: list[APlayer], k_max=10000, with_print=False):
    for k in range(0, k_max):
        state = "---------"
        move = 1
        run = True
        while run:
            for player in players:
                if with_print:
                    print(f"---------------------------------------------------")
                    print(f"- Game {k:06d}, Move {move} -----------------------------")
                    print(f"---------------------------------------------------")
                last_state = state
                state, has_moved = player.move(state)
                if with_print:
                    print(to_state_string(last_state, state))

                if not has_moved:
                    if with_print:
                        print(f"{player.mark}: No more moves")
                    run = False
                if player.has_won(state):
                    if with_print:
                        print(f"{player.mark}: Has won the game")
                    if isinstance(player, MachinePlayer):
                        player.set_value(1.0)
                    run = False

                if not run:
                    break

                move += 1


def choose_player(p1: APlayer, p2: APlayer) -> list[APlayer]:
    # Wer fängt an?
    if uniform() < 0.5:
        result = [p1, p2]
    else:
        result = [p2, p1]

    return result


opponents = choose_player(MachinePlayer(mark='X', with_debug=True), MachinePlayer(mark='O', with_debug=True))
play(opponents, 1000, True)