grid_search_dqn.py

"""
Script que implementa la búsqueda en rejilla de los hiperparámetros que resultan en el mejor agente DDQN.
"""
import json
import os
import random

import gym
import numpy as np
import pandas as pd
import torch

from dqn import DQN, DQNAgent
from playing import play_games_using_agent
from replay_buffer import ExperienceReplayBuffer
from utils import plot_rewards, plot_losses, plot_evaluation_rewards

if __name__ == '__main__':
    # Utilizamos la cpu porque en este caso es más rápida:
    DEVICE = torch.device('cpu')
    # Definimos donde guardar los resultados:
    agent_name = "gs_dqn"
    gs_results_file = f"{agent_name}/experiments.csv"
    os.mkdir(agent_name)

    # Hyperparams:
    MEMORY_SIZE = 10000  # Máxima capacidad del buffer
    BURN_IN = 1000  # Número de pasos iniciales usados para rellenar el buffer antes de entrenar
    MAX_EPISODES = 1000  # Número máximo de episodios (el agente debe aprender antes de llegar a este valor)
    INIT_EPSILON = 1  # Valor inicial de epsilon
    EPSILON_DECAY = .985  # Decaimiento de epsilon
    MIN_EPSILON = 0.01  # Valor mínimo de epsilon en entrenamiento
    GAMMA = 0.99  # Valor gamma de la ecuación de Bellman
    BATCH_SIZES = [16, 32, 64]  # Conjunto a coger del buffer para la red neuronal
    LRS = [0.001, 0.0005, 0.0001]  # Velocidad de aprendizaje
    DNN_UPDS = [1, 3]  # Frecuencia de actualización de la red neuronal
    DNN_SYNCS = [1000, 2000]  # Frecuencia de sincronización de pesos entre la red neuronal y la red objetivo

    # Guardamos todos los hiperparámetros por reproducibilidad:
    all_params = {
        "MEMORY_SIZE": MEMORY_SIZE,
        "BURN_IN": BURN_IN,
        "MAX_EPISODES": MAX_EPISODES,
        "INIT_EPSILON": INIT_EPSILON,
        "EPSILON_DECAY": EPSILON_DECAY,
        "MIN_EPSILON": MIN_EPSILON,
        "GAMMA": GAMMA,
        "BATCH_SIZES": BATCH_SIZES,
        "LRS": LRS,
        "DNN_UPDS": DNN_UPDS,
        "DNN_SYNCS": DNN_SYNCS
    }
    with open(f"{agent_name}/grid.json", "w") as file:
        json.dump(obj=all_params, fp=file)

    # Grid search:
    gs_results = []
    for BATCH_SIZE in BATCH_SIZES:
        for LR in LRS:
            for DNN_UPD in DNN_UPDS:
                for DNN_SYNC in DNN_SYNCS:
                    # Inicialización:
                    env_dict = {'id': 'LunarLander-v2', 'render_mode': 'rgb_array'}
                    environment = gym.make(**env_dict)
                    # Fijamos las semillas utilizadas, por reproducibilidad:
                    RANDOM_SEED = 666
                    random.seed(RANDOM_SEED)
                    torch.manual_seed(RANDOM_SEED)
                    np.random.seed(RANDOM_SEED)
                    environment.reset(seed=RANDOM_SEED)
                    environment.action_space.seed(RANDOM_SEED)

                    # Parameters:
                    parameters = (f"BATCH_SIZE={BATCH_SIZE}, LR={LR}, "
                                  f"DNN_UPD={DNN_UPD}, DNN_SYNC={DNN_SYNC}")
                    print(f"Running experiment with {parameters}.")
                    # Agent initialization:
                    er_buffer = ExperienceReplayBuffer(memory_size=MEMORY_SIZE, burn_in=BURN_IN)
                    dqn = DQN(env=environment, learning_rate=LR, device=DEVICE)
                    dqn_agent = DQNAgent(
                        env=environment,
                        dnnetwork=dqn,
                        buffer=er_buffer,
                        epsilon=INIT_EPSILON,
                        eps_decay=EPSILON_DECAY,
                        batch_size=BATCH_SIZE,
                        min_epsilon=MIN_EPSILON
                    )

                    # Agent training:
                    training_time = dqn_agent.train(
                        gamma=GAMMA,
                        max_episodes=MAX_EPISODES,
                        dnn_update_frequency=DNN_UPD,
                        dnn_sync_frequency=DNN_SYNC
                    )
                    print(f"Training time: {training_time} minutes.")
                    # Training evaluation:
                    plot_rewards(
                        training_rewards=dqn_agent.training_rewards,
                        mean_training_rewards=dqn_agent.mean_training_rewards,
                        reward_threshold=environment.spec.reward_threshold,
                        title=parameters,
                        save_file_name=f'{agent_name}/{parameters}_rewards.png'
                    )
                    plot_losses(
                        training_losses=dqn_agent.training_losses,
                        title=parameters,
                        save_file_name=f'{agent_name}/{parameters}_losses.png'
                    )

                    # Evaluation:
                    eval_eps = 0
                    eval_games_seed = 0
                    tr, _ = play_games_using_agent(
                        environment_dict=env_dict,
                        agent=dqn_agent,
                        n_games=100,
                        eps=eval_eps,
                        games_seed=eval_games_seed
                    )
                    plot_evaluation_rewards(
                        rewards=tr,
                        reward_threshold=environment.spec.reward_threshold,
                        title=parameters,
                        save_file_name=f'{agent_name}/{parameters}_evaluation.png'
                    )
                    # Store metrics:
                    run_results = {
                        'solved': dqn_agent.mean_training_rewards[-1] >= environment.spec.reward_threshold,
                        'train episodes': len(dqn_agent.mean_training_rewards),
                        "training time": training_time,
                        'mean evaluation rewards': round(tr.mean(), 2),
                        'median evaluation rewards': round(np.median(tr), 2),
                        'well landed eval. episodes': sum(tr >= 200),
                        'landed eval. episodes': sum((tr < 200) & (tr >= 100)),
                        'crashed eval. episodes': sum(tr < 100),
                        "LEARNING RATE": LR,
                        "DNN UPD FREQ": DNN_UPD,
                        "DNN SYNC FREQ": DNN_SYNC,
                        "BATCH SIZE": BATCH_SIZE,
                        "MAX EPISODES": MAX_EPISODES
                    }
                    # Update the list:
                    gs_results.append(run_results)
                    # Update the csv:
                    pd.DataFrame(gs_results).to_csv(gs_results_file)