Use Elibitility Trace Actor-Critic to Play Acrobot-v1¶

PyTorch version

In [1]:
%matplotlib inline

import sys
import logging
import copy
import itertools

import numpy as np
np.random.seed(0)
import pandas as pd
import gym
import matplotlib.pyplot as plt
import torch
torch.manual_seed(0)
import torch.autograd as autograd
import torch.nn as nn
import torch.nn.init as init
import torch.optim as optim
import torch.distributions as distributions

logging.basicConfig(level=logging.INFO,
        format='%(asctime)s [%(levelname)s] %(message)s',
        stream=sys.stdout, datefmt='%H:%M:%S')
In [2]:
env = gym.make('Acrobot-v1')
for key in vars(env):
    logging.info('%s: %s', key, vars(env)[key])
for key in vars(env.spec):
    logging.info('%s: %s', key, vars(env.spec)[key])

00:01:37 [INFO] env: <AcrobotEnv<Acrobot-v1>>
00:01:37 [INFO] action_space: Discrete(3)
00:01:37 [INFO] observation_space: Box(-28.274333953857422, 28.274333953857422, (6,), float32)
00:01:37 [INFO] reward_range: (-inf, inf)
00:01:37 [INFO] metadata: {'render.modes': ['human', 'rgb_array'], 'video.frames_per_second': 15}
00:01:37 [INFO] _max_episode_steps: 500
00:01:37 [INFO] _elapsed_steps: None
00:01:37 [INFO] id: Acrobot-v1
00:01:37 [INFO] entry_point: gym.envs.classic_control:AcrobotEnv
00:01:37 [INFO] reward_threshold: -100.0
00:01:37 [INFO] nondeterministic: False
00:01:37 [INFO] max_episode_steps: 500
00:01:37 [INFO] _kwargs: {}
00:01:37 [INFO] _env_name: Acrobot
In [3]:
class ElibilityTraceActorCriticAgent:
    def __init__(self, env):
        self.action_n = env.action_space.n
        self.gamma = 0.99
        self.actor_lambda = 0.9
        self.critic_lambda = 0.9

        self.actor_net = self.build_net(
                input_size=env.observation_space.shape[0],
                hidden_sizes=[100,],
                output_size=env.action_space.n, output_activator=nn.Softmax(1))
        self.actor_optimizer = optim.Adam(self.actor_net.parameters(), 0.0001)
        self.actor_trace = copy.deepcopy(self.actor_net)

        self.critic_net = self.build_net(
                input_size=env.observation_space.shape[0],
                hidden_sizes=[100,], output_size=1)
        self.critic_optimizer = optim.Adam(self.critic_net.parameters(), 0.0002)
        self.critic_loss = nn.MSELoss()
        self.critic_trace = copy.deepcopy(self.critic_net)

    def build_net(self, input_size, hidden_sizes, output_size,
            output_activator=None):
        layers = []
        for input_size, output_size in zip(
                [input_size,] + hidden_sizes, hidden_sizes + [output_size,]):
            layers.append(nn.Linear(input_size, output_size))
            layers.append(nn.ReLU())
        layers = layers[:-1]
        if output_activator:
            layers.append(output_activator)
        net = nn.Sequential(*layers)
        return net

    def reset(self, mode=None):
        self.mode = mode
        if self.mode == 'train':
            self.trajectory = []
            self.discount = 1.

            def weights_init(m):
                if isinstance(m, nn.Linear):
                    init.zeros_(m.weight)
                    init.zeros_(m.bias)
            self.actor_trace.apply(weights_init)
            self.critic_trace.apply(weights_init)

    def step(self, observation, reward, terminated):
        state_tensor = torch.as_tensor(observation, dtype=torch.float).unsqueeze(0)
        prob_tensor = self.actor_net(state_tensor)
        action_tensor = distributions.Categorical(prob_tensor).sample()
        action = action_tensor.numpy()[0]
        if self.mode == 'train':
            self.trajectory += [observation, reward, terminated, action]
            if len(self.trajectory) >= 8:
                self.learn()
            self.discount *= self.gamma
        return action

    def close(self):
        pass

    def update_net(self, target_net, evaluate_net, target_weight, evaluate_weight):
        for target_param, evaluate_param in zip(
                target_net.parameters(), evaluate_net.parameters()):
            target_param.data.copy_(evaluate_weight * evaluate_param.data
                    + target_weight * target_param.data)

    def learn(self):
        state, _, _, action, next_state, reward, terminated, next_action = \
                self.trajectory[-8:]
        state_tensor = torch.as_tensor(state, dtype=torch.float).unsqueeze(0)
        next_state_tensor = torch.as_tensor(next_state, dtype=torch.float).unsqueeze(0)

        pred_tensor = self.critic_net(state_tensor)
        pred = pred_tensor.detach().numpy()[0, 0]
        next_v_tesnor = self.critic_net(next_state_tensor)
        next_v = next_v_tesnor.detach().numpy()[0, 0]
        target = reward + (1. - terminated) * self.gamma * next_v
        td_error = target - pred

        # update actor
        pi_tensor = self.actor_net(state_tensor)[0, action]
        logpi_tensor = torch.log(torch.clamp(pi_tensor, 1e-6, 1.))
        self.actor_optimizer.zero_grad()
        logpi_tensor.backward(retain_graph=True)
        for param, trace in zip(self.actor_net.parameters(),
                self.actor_trace.parameters()):
            trace.data.copy_(self.gamma * self.actor_lambda * trace.data + \
                    self.discount * param.grad)
            param.grad.copy_(-td_error * trace)
        self.actor_optimizer.step()

        # update critic
        v_tensor = self.critic_net(state_tensor)[0, 0]
        self.critic_optimizer.zero_grad()
        v_tensor.backward()
        for param, trace in zip(self.critic_net.parameters(),
                self.critic_trace.parameters()):
            trace.data.copy_(self.gamma * self.critic_lambda * trace.data +
                    param.grad)
            param.grad.copy_(-td_error * trace)
        self.critic_optimizer.step()


agent = ElibilityTraceActorCriticAgent(env)
In [4]:
def play_episode(env, agent, seed=None, mode=None, render=False):
    observation, _ = env.reset(seed=seed)
    reward, terminated, truncated = 0., False, False
    agent.reset(mode=mode)
    episode_reward, elapsed_steps = 0., 0
    while True:
        action = agent.step(observation, reward, terminated)
        if render:
            env.render()
        if terminated or truncated:
            break
        observation, reward, terminated, truncated, _ = env.step(action)
        episode_reward += reward
        elapsed_steps += 1
    agent.close()
    return episode_reward, elapsed_steps


logging.info('==== train ====')
episode_rewards = []
for episode in itertools.count():
    episode_reward, elapsed_steps = play_episode(env, agent, seed=episode,
            mode='train')
    episode_rewards.append(episode_reward)
    logging.info('train episode %d: reward = %.2f, steps = %d',
            episode, episode_reward, elapsed_steps)
    if np.mean(episode_rewards[-10:]) > -120:
        break
plt.plot(episode_rewards)


logging.info('==== test ====')
episode_rewards = []
for episode in range(100):
    episode_reward, elapsed_steps = play_episode(env, agent)
    episode_rewards.append(episode_reward)
    logging.info('test episode %d: reward = %.2f, steps = %d',
            episode, episode_reward, elapsed_steps)
logging.info('average episode reward = %.2f ± %.2f',
        np.mean(episode_rewards), np.std(episode_rewards))

00:01:37 [INFO] ==== train ====
00:01:40 [INFO] train episode 0: reward = -500.00, steps = 500
00:01:43 [INFO] train episode 1: reward = -500.00, steps = 500
00:01:45 [INFO] train episode 2: reward = -500.00, steps = 500
00:01:47 [INFO] train episode 3: reward = -346.00, steps = 347
00:01:48 [INFO] train episode 4: reward = -149.00, steps = 150
00:01:50 [INFO] train episode 5: reward = -311.00, steps = 312
00:01:51 [INFO] train episode 6: reward = -323.00, steps = 324
00:01:52 [INFO] train episode 7: reward = -184.00, steps = 185
00:01:53 [INFO] train episode 8: reward = -143.00, steps = 144
00:01:54 [INFO] train episode 9: reward = -225.00, steps = 226
00:01:56 [INFO] train episode 10: reward = -207.00, steps = 208
00:01:56 [INFO] train episode 11: reward = -171.00, steps = 172
00:01:57 [INFO] train episode 12: reward = -123.00, steps = 124
00:01:58 [INFO] train episode 13: reward = -208.00, steps = 209
00:01:59 [INFO] train episode 14: reward = -188.00, steps = 189
00:02:00 [INFO] train episode 15: reward = -209.00, steps = 210
00:02:01 [INFO] train episode 16: reward = -123.00, steps = 124
00:02:02 [INFO] train episode 17: reward = -158.00, steps = 159
00:02:03 [INFO] train episode 18: reward = -143.00, steps = 144
00:02:04 [INFO] train episode 19: reward = -204.00, steps = 205
00:02:05 [INFO] train episode 20: reward = -165.00, steps = 166
00:02:06 [INFO] train episode 21: reward = -148.00, steps = 149
00:02:07 [INFO] train episode 22: reward = -145.00, steps = 146
00:02:08 [INFO] train episode 23: reward = -115.00, steps = 116
00:02:08 [INFO] train episode 24: reward = -159.00, steps = 160
00:02:09 [INFO] train episode 25: reward = -135.00, steps = 136
00:02:10 [INFO] train episode 26: reward = -117.00, steps = 118
00:02:11 [INFO] train episode 27: reward = -137.00, steps = 138
00:02:11 [INFO] train episode 28: reward = -129.00, steps = 130
00:02:12 [INFO] train episode 29: reward = -102.00, steps = 103
00:02:13 [INFO] train episode 30: reward = -119.00, steps = 120
00:02:13 [INFO] train episode 31: reward = -99.00, steps = 100
00:02:14 [INFO] train episode 32: reward = -122.00, steps = 123
00:02:15 [INFO] train episode 33: reward = -134.00, steps = 135
00:02:15 [INFO] train episode 34: reward = -108.00, steps = 109
00:02:16 [INFO] train episode 35: reward = -131.00, steps = 132
00:02:16 [INFO] ==== test ====
00:02:16 [INFO] test episode 0: reward = -130.00, steps = 131
00:02:16 [INFO] test episode 1: reward = -226.00, steps = 227
00:02:16 [INFO] test episode 2: reward = -133.00, steps = 134
00:02:17 [INFO] test episode 3: reward = -219.00, steps = 220
00:02:17 [INFO] test episode 4: reward = -189.00, steps = 190
00:02:17 [INFO] test episode 5: reward = -106.00, steps = 107
00:02:17 [INFO] test episode 6: reward = -107.00, steps = 108
00:02:17 [INFO] test episode 7: reward = -111.00, steps = 112
00:02:17 [INFO] test episode 8: reward = -162.00, steps = 163
00:02:17 [INFO] test episode 9: reward = -105.00, steps = 106
00:02:17 [INFO] test episode 10: reward = -133.00, steps = 134
00:02:17 [INFO] test episode 11: reward = -123.00, steps = 124
00:02:17 [INFO] test episode 12: reward = -96.00, steps = 97
00:02:18 [INFO] test episode 13: reward = -118.00, steps = 119
00:02:18 [INFO] test episode 14: reward = -105.00, steps = 106
00:02:18 [INFO] test episode 15: reward = -126.00, steps = 127
00:02:18 [INFO] test episode 16: reward = -182.00, steps = 183
00:02:18 [INFO] test episode 17: reward = -119.00, steps = 120
00:02:18 [INFO] test episode 18: reward = -137.00, steps = 138
00:02:18 [INFO] test episode 19: reward = -133.00, steps = 134
00:02:18 [INFO] test episode 20: reward = -287.00, steps = 288
00:02:19 [INFO] test episode 21: reward = -169.00, steps = 170
00:02:19 [INFO] test episode 22: reward = -217.00, steps = 218
00:02:19 [INFO] test episode 23: reward = -272.00, steps = 273
00:02:19 [INFO] test episode 24: reward = -105.00, steps = 106
00:02:19 [INFO] test episode 25: reward = -223.00, steps = 224
00:02:19 [INFO] test episode 26: reward = -142.00, steps = 143
00:02:19 [INFO] test episode 27: reward = -148.00, steps = 149
00:02:20 [INFO] test episode 28: reward = -158.00, steps = 159
00:02:20 [INFO] test episode 29: reward = -144.00, steps = 145
00:02:20 [INFO] test episode 30: reward = -153.00, steps = 154
00:02:20 [INFO] test episode 31: reward = -126.00, steps = 127
00:02:20 [INFO] test episode 32: reward = -210.00, steps = 211
00:02:20 [INFO] test episode 33: reward = -136.00, steps = 137
00:02:20 [INFO] test episode 34: reward = -140.00, steps = 141
00:02:20 [INFO] test episode 35: reward = -123.00, steps = 124
00:02:20 [INFO] test episode 36: reward = -107.00, steps = 108
00:02:21 [INFO] test episode 37: reward = -139.00, steps = 140
00:02:21 [INFO] test episode 38: reward = -118.00, steps = 119
00:02:21 [INFO] test episode 39: reward = -118.00, steps = 119
00:02:21 [INFO] test episode 40: reward = -118.00, steps = 119
00:02:21 [INFO] test episode 41: reward = -250.00, steps = 251
00:02:21 [INFO] test episode 42: reward = -140.00, steps = 141
00:02:21 [INFO] test episode 43: reward = -160.00, steps = 161
00:02:21 [INFO] test episode 44: reward = -115.00, steps = 116
00:02:21 [INFO] test episode 45: reward = -148.00, steps = 149
00:02:22 [INFO] test episode 46: reward = -128.00, steps = 129
00:02:22 [INFO] test episode 47: reward = -206.00, steps = 207
00:02:22 [INFO] test episode 48: reward = -172.00, steps = 173
00:02:22 [INFO] test episode 49: reward = -152.00, steps = 153
00:02:22 [INFO] test episode 50: reward = -142.00, steps = 143
00:02:22 [INFO] test episode 51: reward = -178.00, steps = 179
00:02:22 [INFO] test episode 52: reward = -84.00, steps = 85
00:02:22 [INFO] test episode 53: reward = -115.00, steps = 116
00:02:23 [INFO] test episode 54: reward = -229.00, steps = 230
00:02:23 [INFO] test episode 55: reward = -226.00, steps = 227
00:02:23 [INFO] test episode 56: reward = -104.00, steps = 105
00:02:23 [INFO] test episode 57: reward = -172.00, steps = 173
00:02:23 [INFO] test episode 58: reward = -115.00, steps = 116
00:02:23 [INFO] test episode 59: reward = -141.00, steps = 142
00:02:23 [INFO] test episode 60: reward = -139.00, steps = 140
00:02:23 [INFO] test episode 61: reward = -148.00, steps = 149
00:02:24 [INFO] test episode 62: reward = -256.00, steps = 257
00:02:24 [INFO] test episode 63: reward = -172.00, steps = 173
00:02:24 [INFO] test episode 64: reward = -161.00, steps = 162
00:02:24 [INFO] test episode 65: reward = -134.00, steps = 135
00:02:24 [INFO] test episode 66: reward = -201.00, steps = 202
00:02:24 [INFO] test episode 67: reward = -117.00, steps = 118
00:02:24 [INFO] test episode 68: reward = -125.00, steps = 126
00:02:24 [INFO] test episode 69: reward = -122.00, steps = 123
00:02:24 [INFO] test episode 70: reward = -197.00, steps = 198
00:02:25 [INFO] test episode 71: reward = -145.00, steps = 146
00:02:25 [INFO] test episode 72: reward = -124.00, steps = 125
00:02:25 [INFO] test episode 73: reward = -147.00, steps = 148
00:02:25 [INFO] test episode 74: reward = -134.00, steps = 135
00:02:25 [INFO] test episode 75: reward = -129.00, steps = 130
00:02:25 [INFO] test episode 76: reward = -107.00, steps = 108
00:02:25 [INFO] test episode 77: reward = -171.00, steps = 172
00:02:25 [INFO] test episode 78: reward = -82.00, steps = 83
00:02:25 [INFO] test episode 79: reward = -117.00, steps = 118
00:02:26 [INFO] test episode 80: reward = -131.00, steps = 132
00:02:26 [INFO] test episode 81: reward = -145.00, steps = 146
00:02:26 [INFO] test episode 82: reward = -111.00, steps = 112
00:02:26 [INFO] test episode 83: reward = -170.00, steps = 171
00:02:26 [INFO] test episode 84: reward = -161.00, steps = 162
00:02:26 [INFO] test episode 85: reward = -109.00, steps = 110
00:02:26 [INFO] test episode 86: reward = -111.00, steps = 112
00:02:26 [INFO] test episode 87: reward = -98.00, steps = 99
00:02:26 [INFO] test episode 88: reward = -142.00, steps = 143
00:02:26 [INFO] test episode 89: reward = -155.00, steps = 156
00:02:27 [INFO] test episode 90: reward = -85.00, steps = 86
00:02:27 [INFO] test episode 91: reward = -101.00, steps = 102
00:02:27 [INFO] test episode 92: reward = -165.00, steps = 166
00:02:27 [INFO] test episode 93: reward = -125.00, steps = 126
00:02:27 [INFO] test episode 94: reward = -114.00, steps = 115
00:02:27 [INFO] test episode 95: reward = -145.00, steps = 146
00:02:27 [INFO] test episode 96: reward = -131.00, steps = 132
00:02:27 [INFO] test episode 97: reward = -126.00, steps = 127
00:02:27 [INFO] test episode 98: reward = -173.00, steps = 174
00:02:27 [INFO] test episode 99: reward = -112.00, steps = 113
00:02:27 [INFO] average episode reward = -146.28 ± 41.48
In [5]:
env.close()