utils.py

from mcts import MCTS
import numpy as np
from kaggle_environments import make, evaluate
import torch

class KaggleAgent:
    def __init__(self, model, game, args):
        self.model = model
        self.game = game
        self.args = args
        if self.args['search']:
            self.mcts = MCTS(model, game, args)

    def run(self, obs, conf):
        player = obs['mark'] if obs['mark'] == 1 else -1
        state = np.array(obs['board']).reshape(self.game.row_count, self.game.column_count)
        state[state==2] = -1
        
        state = self.game.change_perspective(state, player)        

        if self.args['search']:
            policy = self.mcts.search(state)

        else:
            policy, _ = self.model(
                torch.tensor(self.game.get_encoded_state(state), device=self.model.device).unsqueeze(0)
            )

        valid_moves = self.game.get_valid_moves(state)
        policy *= valid_moves
        policy /= np.sum(policy)

        if self.args['temperature'] == 0:
            action = int(np.argmax(policy))
        elif self.args['temperature'] == float('inf'):
            action = np.random.choice([r for r in range(self.game.action_size) if policy[r] > 0])
        else:
            policy = policy ** (1 / self.args['temperature'])
            policy /= np.sum(policy)
            action = np.random.choice(self.game.action_size, p=policy)

        return action

def evaluateKaggle(gameName, players, num_iterations=1):
    if num_iterations == 1:
        env = make(gameName, debug=True)
        env.run(players)
        return env.render(mode="ipython")

    results = np.array(evaluate(gameName, players, num_episodes=num_iterations))[:, 0]
    print(f"""
Player 1 | Wins: {np.sum(results == 1)} | Draws: {np.sum(results == 0)} | Losses: {np.sum(results == -1)}
Player 2 | Wins: {np.sum(results == -1)} | Draws: {np.sum(results == 0)} | Losses: {np.sum(results == 1)}
    """)