def train(n_epoch):
env = Environment()
arena = Arena()
print('Making Models...')
agents = [None, get_ddqn(env), get_ddqn(env)]
agents[1].load('./models/agent1')
# agents[1].replay_buffer.load('./models/replay1.npz')
agents[2].load('./models/agent2')
# agents[2].replay_buffer.load('./models/replay2.npz')
for e in range(n_epoch):
print('-' * 30)
print(f'{e} Epoch Start...')
print('-' * 30)
for i in range(100):
obs = env.reset(1)
done = False
rewards = [0] * 3
value = 0
last = None
while not done:
last = env.state
action = agents[env.current_player].act_and_train(
obs, rewards[env.current_player])
obs, r, done, value = env.step(action)
rewards[env.current_player - 1] = r
assert last is not None
if value == 1:
agents[1].stop_episode_and_train(env.state, 1, True)
agents[-1].stop_episode_and_train(last, -1, True)
elif value == -1:
agents[-1].stop_episode_and_train(env.state, 1, True)
agents[1].stop_episode_and_train(last, -1, True)
elif value == 0:
agents[1].stop_episode_and_train(env.state, 0, True)
agents[-1].stop_episode_and_train(env.state, 0, True)
elif value == 2:
agents[1].stop_episode_and_train(env.state, -1, True)
agents[-1].stop_episode()
elif value == -2:
agents[-1].stop_episode_and_train(env.state, -1, True)
agents[1].stop_episode()
# print(env.print_board())
# print('Result: ', value)
print(f'{i} Episode Ended')
res1, res2 = arena.battles(agents[1], agents[2])
if res1:
dir_name = 'models/' + 'agent1'
print('Saving model', dir_name)
agents[1].save(dir_name)
agents[1].replay_buffer.save('models/replay1.npz')
if res2:
dir_name = 'models/' + 'agent2'
print('Saving model', dir_name)
agents[2].save(dir_name)
agents[2].replay_buffer.save('models/replay2.npz')
