Use Closed-Form Policy to Play LunarLander-v2¶

In [1]:
import sys
import logging
import itertools

import numpy as np
np.random.seed(0)
import gym

logging.basicConfig(level=logging.INFO,
        format='%(asctime)s [%(levelname)s] %(message)s',
        stream=sys.stdout, datefmt='%H:%M:%S')
In [2]:
env = gym.make('LunarLander-v2')
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])

00:00:00 [INFO] env: <LunarLander<LunarLander-v2>>
00:00:00 [INFO] action_space: Discrete(4)
00:00:00 [INFO] observation_space: Box(-inf, inf, (8,), float32)
00:00:00 [INFO] reward_range: (-inf, inf)
00:00:00 [INFO] metadata: {'render.modes': ['human', 'rgb_array'], 'video.frames_per_second': 50}
00:00:00 [INFO] _max_episode_steps: 1000
00:00:00 [INFO] _elapsed_steps: None
00:00:00 [INFO] id: LunarLander-v2
00:00:00 [INFO] entry_point: gym.envs.box2d:LunarLander
00:00:00 [INFO] reward_threshold: 200
00:00:00 [INFO] nondeterministic: False
00:00:00 [INFO] max_episode_steps: 1000
00:00:00 [INFO] _kwargs: {}
00:00:00 [INFO] _env_name: LunarLander
In [3]:
class ClosedFormAgent:
    def __init__(self, _):
        pass

    def reset(self, mode=None):
        pass

    def step(self, observation, reward, terminated):
        x, y, v_x, v_y, angle, v_angle, contact_left, contact_right = observation

        if contact_left or contact_right:  # legs have contact
            f_y = -10. * v_y - 1.
            f_angle = 0.
        else:
            f_y = 5.5 * np.abs(x) - 10. * y - 10. * v_y - 1.
            f_angle = -np.clip(5. * x + 10. * v_x, -4, 4) + 10. * angle + 20. \
                    * v_angle

        if np.abs(f_angle) <= 1 and f_y <= 0:
            action = 0 # do nothing
        elif np.abs(f_angle) < f_y:
            action = 2 # main engine
        elif f_angle < 0.:
            action = 1 # left engine
        else:
            action = 3 # right engine
        return action

    def close(self):
        pass


agent = ClosedFormAgent(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('==== 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))

00:00:00 [INFO] ==== test ====
00:00:00 [INFO] test episode 0: reward = 251.10, steps = 175
00:00:00 [INFO] test episode 1: reward = 263.86, steps = 158
00:00:00 [INFO] test episode 2: reward = 296.23, steps = 228
00:00:00 [INFO] test episode 3: reward = 290.20, steps = 201
00:00:00 [INFO] test episode 4: reward = 309.30, steps = 199
00:00:00 [INFO] test episode 5: reward = 304.04, steps = 171
00:00:00 [INFO] test episode 6: reward = 260.22, steps = 199
00:00:00 [INFO] test episode 7: reward = 277.66, steps = 170
00:00:00 [INFO] test episode 8: reward = 255.32, steps = 181
00:00:00 [INFO] test episode 9: reward = 261.03, steps = 168
00:00:00 [INFO] test episode 10: reward = 296.68, steps = 208
00:00:00 [INFO] test episode 11: reward = 300.02, steps = 222
00:00:00 [INFO] test episode 12: reward = 303.89, steps = 234
00:00:00 [INFO] test episode 13: reward = 255.41, steps = 178
00:00:00 [INFO] test episode 14: reward = 267.43, steps = 180
00:00:00 [INFO] test episode 15: reward = 309.62, steps = 264
00:00:01 [INFO] test episode 16: reward = 264.02, steps = 214
00:00:01 [INFO] test episode 17: reward = 297.51, steps = 216
00:00:01 [INFO] test episode 18: reward = 261.13, steps = 195
00:00:01 [INFO] test episode 19: reward = 260.21, steps = 204
00:00:01 [INFO] test episode 20: reward = 266.38, steps = 248
00:00:01 [INFO] test episode 21: reward = 253.89, steps = 210
00:00:01 [INFO] test episode 22: reward = 254.27, steps = 186
00:00:01 [INFO] test episode 23: reward = 198.83, steps = 604
00:00:01 [INFO] test episode 24: reward = 289.12, steps = 225
00:00:01 [INFO] test episode 25: reward = 272.47, steps = 191
00:00:01 [INFO] test episode 26: reward = 316.36, steps = 267
00:00:02 [INFO] test episode 27: reward = 235.75, steps = 451
00:00:02 [INFO] test episode 28: reward = 289.84, steps = 211
00:00:02 [INFO] test episode 29: reward = 282.75, steps = 194
00:00:02 [INFO] test episode 30: reward = 249.47, steps = 199
00:00:02 [INFO] test episode 31: reward = 257.96, steps = 304
00:00:02 [INFO] test episode 32: reward = 246.41, steps = 243
00:00:02 [INFO] test episode 33: reward = 285.99, steps = 228
00:00:02 [INFO] test episode 34: reward = 239.66, steps = 165
00:00:02 [INFO] test episode 35: reward = 283.29, steps = 210
00:00:02 [INFO] test episode 36: reward = 244.06, steps = 170
00:00:02 [INFO] test episode 37: reward = 255.37, steps = 208
00:00:02 [INFO] test episode 38: reward = 261.29, steps = 171
00:00:02 [INFO] test episode 39: reward = 253.82, steps = 176
00:00:02 [INFO] test episode 40: reward = 302.92, steps = 237
00:00:02 [INFO] test episode 41: reward = 260.70, steps = 167
00:00:02 [INFO] test episode 42: reward = 255.26, steps = 199
00:00:03 [INFO] test episode 43: reward = 237.06, steps = 179
00:00:03 [INFO] test episode 44: reward = 283.01, steps = 148
00:00:03 [INFO] test episode 45: reward = 258.90, steps = 183
00:00:03 [INFO] test episode 46: reward = 294.71, steps = 251
00:00:03 [INFO] test episode 47: reward = 284.71, steps = 161
00:00:03 [INFO] test episode 48: reward = 281.09, steps = 251
00:00:03 [INFO] test episode 49: reward = 238.25, steps = 192
00:00:03 [INFO] test episode 50: reward = 301.94, steps = 240
00:00:03 [INFO] test episode 51: reward = 263.24, steps = 175
00:00:03 [INFO] test episode 52: reward = 260.24, steps = 173
00:00:03 [INFO] test episode 53: reward = 309.02, steps = 196
00:00:03 [INFO] test episode 54: reward = 294.00, steps = 167
00:00:03 [INFO] test episode 55: reward = 246.90, steps = 180
00:00:03 [INFO] test episode 56: reward = 262.93, steps = 215
00:00:03 [INFO] test episode 57: reward = 285.19, steps = 206
00:00:03 [INFO] test episode 58: reward = 263.72, steps = 193
00:00:03 [INFO] test episode 59: reward = -7.60, steps = 102
00:00:03 [INFO] test episode 60: reward = 260.06, steps = 177
00:00:03 [INFO] test episode 61: reward = 280.44, steps = 211
00:00:03 [INFO] test episode 62: reward = 266.07, steps = 175
00:00:03 [INFO] test episode 63: reward = 272.33, steps = 155
00:00:04 [INFO] test episode 64: reward = 302.49, steps = 205
00:00:04 [INFO] test episode 65: reward = 293.76, steps = 227
00:00:04 [INFO] test episode 66: reward = 280.02, steps = 163
00:00:04 [INFO] test episode 67: reward = 282.13, steps = 202
00:00:04 [INFO] test episode 68: reward = -215.60, steps = 308
00:00:04 [INFO] test episode 69: reward = 306.33, steps = 263
00:00:04 [INFO] test episode 70: reward = 290.63, steps = 151
00:00:04 [INFO] test episode 71: reward = 290.89, steps = 177
00:00:04 [INFO] test episode 72: reward = 271.69, steps = 207
00:00:04 [INFO] test episode 73: reward = 266.11, steps = 161
00:00:04 [INFO] test episode 74: reward = 284.49, steps = 146
00:00:04 [INFO] test episode 75: reward = 266.26, steps = 242
00:00:04 [INFO] test episode 76: reward = 278.96, steps = 363
00:00:05 [INFO] test episode 77: reward = 270.56, steps = 231
00:00:05 [INFO] test episode 78: reward = 269.81, steps = 215
00:00:05 [INFO] test episode 79: reward = 293.34, steps = 222
00:00:05 [INFO] test episode 80: reward = 261.76, steps = 191
00:00:05 [INFO] test episode 81: reward = 277.13, steps = 221
00:00:05 [INFO] test episode 82: reward = 306.04, steps = 187
00:00:05 [INFO] test episode 83: reward = 276.49, steps = 219
00:00:05 [INFO] test episode 84: reward = 284.92, steps = 220
00:00:05 [INFO] test episode 85: reward = 268.83, steps = 187
00:00:05 [INFO] test episode 86: reward = 287.16, steps = 178
00:00:05 [INFO] test episode 87: reward = 279.68, steps = 163
00:00:05 [INFO] test episode 88: reward = 296.61, steps = 238
00:00:05 [INFO] test episode 89: reward = 285.85, steps = 154
00:00:05 [INFO] test episode 90: reward = 280.37, steps = 217
00:00:05 [INFO] test episode 91: reward = 276.90, steps = 181
00:00:05 [INFO] test episode 92: reward = 216.17, steps = 291
00:00:06 [INFO] test episode 93: reward = 19.45, steps = 1000
00:00:06 [INFO] test episode 94: reward = 259.90, steps = 169
00:00:06 [INFO] test episode 95: reward = 262.98, steps = 336
00:00:06 [INFO] test episode 96: reward = 257.21, steps = 259
00:00:06 [INFO] test episode 97: reward = 239.49, steps = 175
00:00:06 [INFO] test episode 98: reward = 280.35, steps = 237
00:00:06 [INFO] test episode 99: reward = 245.76, steps = 194
00:00:06 [INFO] average episode reward = 262.72 ± 64.50
In [5]:
env.close()