backgammon/game.py

from board import Board
from player import Player
from bot import Bot
from restore_bot import RestoreBot
from cup import Cup
from eval import Eval

import numpy as np
import sys
import time

class Game:
    
    def __init__(self, config = None):
        self.config = config
        self.board = Board.initial_state

        self.p1 = None
        self.p2 = None
        # TODO remove this
        self.cup = Cup()

    def set_up_bots(self):
        self.p1 = Bot(1, config = self.config)
        self.p2 = Bot(1, config = self.config)
        
    def roll(self):
        return self.cup.roll()

    def best_move_and_score(self):
        roll = self.roll()
        move_and_val = self.p1.make_move(self.board, self.p1.get_sym(), roll)
        self.board = move_and_val[0]
        return move_and_val

    def next_round(self):
        roll = self.roll()
        #print(roll)
        self.board = Board.flip(self.p2.make_move(Board.flip(self.board), self.p2.get_sym(), roll)[0])
        return self.board

    def board_state(self):
        return self.board

    def play_against_player(self):
        self.board = Board.initial_state
        coin_flip = random.random()
        if coin_flip > 0.5:
            user_color = input("Pick a number, 1 (white) or -1 (black)")
            if int(user_color) == 1:
                p1 = Player(1)
                p2 = Bot(-1)
            else:
                p1 = Bot(1)
                p2 = Player(-1)
        else:
            p1 = Bot(1)
            p2 = Player(-1)

        # Since we have to make sure that the Bot always plays as if it's white, we have to flip
        # the board when it's not actually.
        if p1.__name__ == "Bot" and p1.get_sym() == 1:
            while Board.outcome(self.board) == None:
                roll = self.roll()
                self.board = p1.make_move(self.board, p1.get_sym(), roll)
                roll = self.roll()
                self.board = p2.make_move(self.board, p2.get_sym(), roll)
        if p1.__name__ == "Bot" and p1.get_sym() == -1:
            while Board.outcome(self.board) == None:
                roll = self.roll()
                self.board = Board.flip(p1.make_move(Board.flip(self.board), p1.get_sym(), roll))
                roll = self.roll()
                self.board = p2.make_move(self.board, p2.get_sym(), roll)
        if p2.__name__ == "Bot" and p1.get_sym() == -1:
            while Board.outcome(self.board) == None:
                roll = self.roll()
                self.board = p1.make_move(self.board, p1.get_sym(), roll)
                roll = self.roll()
                self.board = Board.flip(p2.make_move(Board.flip(self.board), p2.get_sym(), roll))
        
            
        print(Board.outcome(self.board))
            

    def train_model(self, episodes=1000, save_step_size = 100, trained_eps = 0):
        start_time = time.time()

        def print_time_estimate(eps_completed):
            cur_time      = time.time()
            time_diff     = cur_time - start_time
            eps_per_sec   = eps_completed / time_diff
            secs_per_ep   = time_diff / eps_completed
            eps_remaining = (episodes - eps_completed)
            sys.stderr.write("[TRAIN] Averaging {per_sec} episodes per second\n".format(per_sec = round(eps_per_sec, 2)))
            sys.stderr.write("[TRAIN] {eps_remaining} episodes remaining; approx. {time_remaining} seconds remaining\n".format(eps_remaining = eps_remaining, time_remaining = int(eps_remaining * secs_per_ep)))

        
        sys.stderr.write("[TRAIN] Training {} episodes and save_step_size {}\n".format(episodes, save_step_size))
        outcomes = []
        for episode in range(1, episodes + 1):
            sys.stderr.write("[TRAIN] Episode {}".format(episode + trained_eps))
            self.board = Board.initial_state

            prev_board, prev_board_value = self.best_move_and_score()
            # find the best move here, make this move, then change turn as the
            # first thing inside of the while loop and then call
            # best_move_and_score to get V_t+1

            while Board.outcome(self.board) is None:
                self.next_round()
                cur_board, cur_board_value = self.best_move_and_score()
                self.p1.get_network().train(prev_board, cur_board_value)
                prev_board = cur_board

                # print("-"*30)
                # print(Board.pretty(self.board))
                # print("/"*30)
            sys.stderr.write("\t outcome {}".format(Board.outcome(self.board)[1]))
            outcomes.append(Board.outcome(self.board)[1])
            final_score = np.array([ Board.outcome(self.board)[1] ]).reshape((1, 1))
            self.p1.get_network().train(prev_board, final_score)

            sys.stderr.write("\n")
            
            if episode % min(save_step_size, episodes) == 0:
                sys.stderr.write("[TRAIN] Saving model...\n")
                self.p1.get_network().save_model(episode+trained_eps)
                sys.stderr.write("[TRAIN] Loading model for training opponent...\n")
                self.p2.restore_model()

            if episode % 50 == 0:
                print_time_estimate(episode)

        sys.stderr.write("[TRAIN] Saving model for final episode...\n")
        self.p1.get_network().save_model(episode+trained_eps)
        self.p2.restore_model()
        
        return outcomes

    def next_round_test(self):
        print(self.board)
        print()
        self.next_round()
        print("--------------------------------")
        print(self.board)
        print("--------------------------------")

    def eval(self, trained_eps = 0):
        def do_eval(method, episodes = 1000, trained_eps = 0):
            start_time = time.time()

            def print_time_estimate(eps_completed):
                cur_time      = time.time()
                time_diff     = cur_time - start_time
                eps_per_sec   = eps_completed / time_diff
                secs_per_ep   = time_diff / eps_completed
                eps_remaining = (episodes - eps_completed)
                sys.stderr.write("[EVAL ] Averaging {per_sec} episodes per second\n".format(per_sec = round(eps_per_sec, 2)))
                sys.stderr.write("[EVAL ] {eps_remaining} episodes remaining; approx. {time_remaining} seconds remaining\n".format(eps_remaining = eps_remaining, time_remaining = int(eps_remaining * secs_per_ep)))

            sys.stderr.write("[EVAL ] Evaluating {eps} episode(s) with method '{method}'\n".format(eps=episodes, method=method))
            
            if method == 'random':
                outcomes = []
                for i in range(1, episodes + 1):
                    sys.stderr.write("[EVAL ] Episode {}".format(i))
                    self.board = Board.initial_state
                    while Board.outcome(self.board) is None:
                        roll = self.roll()
                        self.board = (self.p1.make_move(self.board, self.p1.get_sym(), roll))[0]
                        roll = self.roll()
                        self.board = Board.flip(Eval.make_random_move(Board.flip(self.board), self.p2.get_sym(), roll))
                    sys.stderr.write("\t outcome {}".format(Board.outcome(self.board)[1]))
                    outcomes.append(Board.outcome(self.board)[1])
                    sys.stderr.write("\n")

                    if i % 50 == 0:
                        print_time_estimate(i)
                return outcomes
            elif method == 'pubeval':
                outcomes = []
                # Add the evaluation code for pubeval, the bot has a method make_pubeval_move(board, sym, roll), which can be used to get the best move according to pubeval
                for i in range(1, episodes + 1):
                    sys.stderr.write("[EVAL ] Episode {}".format(i))
                    self.board = Board.initial_state
                    while Board.outcome(self.board) is None:
                        roll = self.roll()
                        self.board = (self.p1.make_move(self.board, self.p1.get_sym(), roll))[0]
                        roll = self.roll()
                        self.board = Board.flip(Eval.make_pubeval_move(self.board, self.p2.get_sym(), roll)[0][0:26])
                    sys.stderr.write("\t outcome {}".format(Board.outcome(self.board)[1]))
                    outcomes.append(Board.outcome(self.board)[1])
                    sys.stderr.write("\n")

                    if i % 10 == 0:
                        print_time_estimate(i)
                    
                return outcomes
            else:
                sys.stderr.write("[EVAL ] Evaluation method '{}' is not defined\n".format(method))
                return [0]
            
        return [ (method, do_eval(method,
                                  self.config['episode_count'],
                                  trained_eps = trained_eps))
                 for method
                 in self.config['eval_methods'] ]

    def play(self, episodes = 1000):
        outcomes = []
        for i in range(episodes):

            self.board = Board.initial_state
            while Board.outcome(self.board) is None:
               # count += 1
           #     print("Turn:",count)

                roll = self.roll()

#                print("type of board: ", type(self.board))
#                print("Board:",self.board)
#                print("{} rolled: {}".format(self.p1.get_sym(), roll))

                self.board = (self.p1.make_random_move(self.board, self.p1.get_sym(), roll))
                
    #            print(self.board)

#                print()
            
#                count += 1
            
                roll = self.roll()
 #               print("{} rolled: {}".format(self.p2.get_sym(), roll))
                self.board = Board.flip(self.p2.make_random_move(Board.flip(self.board), self.p2.get_sym(), roll))


            if Board.outcome(self.board)[1] > 0:
                print_winner = "1: White, " + str(Board.outcome(self.board))
            else:
                print_winner = "-1: Black " + str(Board.outcome(self.board))
            outcomes.append(Board.outcome(self.board)[1])
            print("The winner is {}!".format(print_winner))
            print("Round:",i)
#            print("Final board:",Board.pretty(self.board))
        return outcomes
#            return count

highest = 0