from train_cartpole import run_episode
from agent.networks import MLP
import numpy as np
np.random.seed(0)
if __name__ == "__main__":
env = gym.make("CartPole-v0").unwrapped
# TODO: load DQN agent
# ...
states_dim = 4
action_dim = 2
Q = MLP(states_dim, action_dim)
Q_target = MLP(states_dim, action_dim)
agent = DQNAgent(Q, Q_target, action_dim, double=True)
agent.load("./models_cartpole/dqn_agent_fixed_1.pt")
n_test_episodes = 15
episode_rewards = []
for i in range(n_test_episodes):
stats = run_episode(env,
agent,
deterministic=True,
do_training=False,
rendering=True)
episode_rewards.append(stats.episode_reward)
# save results in a dictionary and write them into a .json file
# Hint: CartPole is considered solved when the average reward is greater than or equal to 195.0 over 100 consecutive trials.
env = gym.make("CartPole-v0").unwrapped
#import pdb; pdb.set_trace()
state_dim = 4
num_actions = 2
# TODO:
# 1. init Q network and target network (see dqn/networks.py)
# 2. init DQNAgent (see dqn/dqn_agent.py)
# 3. train DQN agent with train_online(...)
# Duelling DQN or Not
Duel = False
num_episodes = 2000
if Duel:
Q = MLP_Duel(state_dim, num_actions)
Q_target = MLP_Duel(state_dim, num_actions)
else:
Q = MLP(state_dim, num_actions)
Q_target = MLP(state_dim, num_actions)
DQNAgent = DQNAgent(Q,
Q_target,
num_actions,
double=True,
history_length=1e6)
train_online(env, DQNAgent, num_episodes, epsilon_decay=False)
if __name__ == "__main__":
# You find information about cartpole in
# https://github.com/openai/gym/wiki/CartPole-v0
# Hint: CartPole is considered solved when the average reward is greater than or equal to 195.0 over 100 consecutive trials.
parser = argparse.ArgumentParser()
parser.add_argument("-i", "--interrupt", action='store_true', help="Save model if interrupted",
default=False, required=False)
parser.add_argument('-e', "--episodes", type=int, help="num episodes to try", default=500, required=False)
parser.add_argument('-s', "--steps", type=int, help="num steps per episode", default=200, required=False)
parser.add_argument("-r", "--render", action='store_true', help="render during training and evaluation",
default=False, required=False)
args = parser.parse_args()
print(args)
env = gym.make("CartPole-v0").unwrapped
state_dim = 4
num_actions = 2
# TODO:
# 1. init Q network and target network (see dqn/networks.py)
Q_network = MLP(state_dim, num_actions)
Q_target = MLP(state_dim, num_actions)
# 2. init DQNAgent (see dqn/dqn_agent.py)
agent = DQNAgent(Q=Q_network, Q_target=Q_target, num_actions=num_actions, buffer_size=1e5, lr=1e-4)
# 3. train DQN agent with train_online(...)
train_online(env=env, agent=agent, num_episodes=args.episodes, max_timesteps=args.steps,
eval_cycle=20, num_eval_episodes=5, rendering=args.render,
tensorboard_dir='./tensorboard', save_interrupt=args.interrupt)
import json
from agent.dqn_agent import DQNAgent
from train_cartpole import run_episode
from agent.networks import *
import numpy as np
from agent.networks import MLP
import torch
np.random.seed(0)
if __name__ == "__main__":
env = gym.make("CartPole-v0").unwrapped
# TODO: load DQN agent
# ...
Q = MLP(state_dim = 4,action_dim = 2)
Q_target = MLP(state_dim = 4, action_dim = 2)
agent = DQNAgent(Q, Q_target, num_actions = 2)
agent.load("./models_cartpole/dqn_agent-perfect.pt")
n_test_episodes = 15
episode_rewards = []
for i in range(n_test_episodes):
stats = run_episode(env, agent, deterministic=True, do_training=False, rendering=True,epsilon=0)
episode_rewards.append(stats.episode_reward)
# save results in a dictionary and write them into a .json file
results = dict()
results["episode_rewards"] = episode_rewards
results["mean"] = np.array(episode_rewards).mean()
results["std"] = np.array(episode_rewards).std()