tran.py

import tensorflow.compat.v1 as tf
from open_spiel.python.algorithms import dqn
from open_spiel.python import rl_environment
import random

tf.disable_v2_behavior()

game = "laser_tag"
num_players = 2

# env_configs = {"columns": 5, "rows": 5}
env = rl_environment.Environment("laser_tag")
num_actions = env.action_spec()["num_actions"]
num_train_episodes = 10

checkpoint_dir = '/Users/shawnsun/Desktop/CS486/Versions/models/dqn10'



# Set DQN hyperparameters
hidden_layers_sizes = 64
replay_buffer_capacity = int(1e5)
batch_size = 32
num_train_episodes = int(1)

# Train the DQN model
with tf.Session() as sess:
    model = dqn.DQN(
        session=sess,
        player_id=0,
        state_representation_size=env.observation_spec()["info_state"][0],
        num_actions=num_actions,
        hidden_layers_sizes=hidden_layers_sizes,
        replay_buffer_capacity=replay_buffer_capacity,
        batch_size=batch_size)
    sess.run(tf.global_variables_initializer())

    for ep in range(num_train_episodes):
        time_step = env.reset()
        while not time_step.last():
            action = model.step(time_step).action
            # legal_actions = time_step.observation
            time_step = env.step([action, random.choice(range(num_actions))])

        model.step(time_step)

    model_variables = tf.trainable_variables()
    for var in model_variables:
        print(var)
    model.save(checkpoint_dir)