RL: my_learning_agent.py

Keywords:

cross-entropy method、noisy_evaluation、BinaryActionLinearPolicy、do_rollout、mean、std、

_policies.py

# Support code for cem.py

class BinaryActionLinearPolicy(object):
    def __init__(self, theta):
        self.w = theta[:-1]
        self.b = theta[-1]
    def act(self, ob):
        y = ob.dot(self.w) + self.b
        a = int(y < 0)
        return a

class ContinuousActionLinearPolicy(object):
    def __init__(self, theta, n_in, n_out):
        assert len(theta) == (n_in + 1) * n_out
        self.W = theta[0 : n_in * n_out].reshape(n_in, n_out)
        self.b = theta[n_in * n_out : None].reshape(1, n_out)
    def act(self, ob):
        a = ob.dot(self.W) + self.b
        return a

my_learning_agent.py

from __future__ import print_function

import gym
from gym import wrappers, logger
import numpy as np
from six.moves import cPickle as pickle
import json, sys, os
from os import path
from _policies import BinaryActionLinearPolicy # Different file so it can be unpickled
import argparse

def cem(f, th_mean, batch_size, n_iter, elite_frac, initial_std=1.0):
    """
    Generic implementation of the cross-entropy method for maximizing a black-box function
    f: a function mapping from vector -> scalar
    th_mean: initial mean over input distribution
    batch_size: number of samples of theta to evaluate per batch
    n_iter: number of batches
    elite_frac: each batch, select this fraction of the top-performing samples
    initial_std: initial standard deviation over parameter vectors
    """
    n_elite = int(np.round(batch_size*elite_frac))
    th_std = np.ones_like(th_mean) * initial_std

    for _ in range(n_iter):
        ths = np.array([th_mean + dth for dth in  th_std[None,:]*np.random.randn(batch_size, th_mean.size)])
        ys = np.array([f(th) for th in ths])
        elite_inds = ys.argsort()[::-1][:n_elite]
        elite_ths = ths[elite_inds]
        th_mean = elite_ths.mean(axis=0)
        th_std = elite_ths.std(axis=0)
        yield {'ys' : ys, 'theta_mean' : th_mean, 'y_mean' : ys.mean()}

def do_rollout(agent, env, num_steps, render=False):
    total_rew = 0
    ob = env.reset()
    for t in range(num_steps):
        a = agent.act(ob)
        (ob, reward, done, _info) = env.step(a)
        total_rew += reward
        if render and t%3==0: env.render()
        if done: break
    return total_rew, t+1

if __name__ == '__main__':
    logger.set_level(logger.INFO)

    parser = argparse.ArgumentParser()
    parser.add_argument('--display', action='store_true')
    parser.add_argument('target', nargs="?", default="CartPole-v0")
    args = parser.parse_args()

    env = gym.make(args.target)
    env.seed(0)
    np.random.seed(0)
    params = dict(n_iter=100, batch_size=10, elite_frac = 0.2)
    num_steps = 200

    def noisy_evaluation(theta):
        agent = BinaryActionLinearPolicy(theta)
        rew, T = do_rollout(agent, env, num_steps)
        return rew

    # Train the agent, and snapshot each stage
    for (i, iterdata) in enumerate(cem(noisy_evaluation, np.zeros(env.observation_space.shape[0]+1), **params)):
        print('Iteration %2i. Episode mean reward: %7.3f'%(i, iterdata['y_mean']))
        agent = BinaryActionLinearPolicy(iterdata['theta_mean'])
        do_rollout(agent, env, 200, render=True)


    env.close()


# Results:
python my_learning_agent.py CartPole-v0

INFO: Making new env: CartPole-v0
Iteration  0. Episode mean reward:  23.800
Iteration  1. Episode mean reward:  92.000
Iteration  2. Episode mean reward: 158.400
Iteration  3. Episode mean reward: 179.100
Iteration  4. Episode mean reward: 186.000
Iteration  5. Episode mean reward: 188.300
Iteration  6. Episode mean reward: 180.900
Iteration  7. Episode mean reward: 188.700
Iteration  8. Episode mean reward: 188.600
Iteration  9. Episode mean reward: 185.300
Iteration 10. Episode mean reward: 191.900
Iteration 11. Episode mean reward: 193.000
Iteration 12. Episode mean reward: 188.300
Iteration 13. Episode mean reward: 183.400
Iteration 14. Episode mean reward: 180.400
Iteration 15. Episode mean reward: 197.100
Iteration 16. Episode mean reward: 193.200
Iteration 17. Episode mean reward: 188.500
Iteration 18. Episode mean reward: 182.800
Iteration 19. Episode mean reward: 193.900
......
......
......
Iteration 81. Episode mean reward: 175.400
Iteration 82. Episode mean reward: 183.500
Iteration 83. Episode mean reward: 195.800
Iteration 84. Episode mean reward: 191.300
Iteration 85. Episode mean reward: 192.000
Iteration 86. Episode mean reward: 196.300
Iteration 87. Episode mean reward: 197.300
Iteration 88. Episode mean reward: 184.000
Iteration 89. Episode mean reward: 192.800
Iteration 90. Episode mean reward: 184.000
Iteration 91. Episode mean reward: 184.700
Iteration 92. Episode mean reward: 184.500
Iteration 93. Episode mean reward: 192.400
Iteration 94. Episode mean reward: 196.000
Iteration 95. Episode mean reward: 193.800
Iteration 96. Episode mean reward: 183.400
Iteration 97. Episode mean reward: 196.400
Iteration 98. Episode mean reward: 192.400
Iteration 99. Episode mean reward: 178.500

At a glance: