QLearningFrozenLakeSolution2.py

import numpy as np
import scipy
import gym
from collections import defaultdict
from gym import wrappers


GAMMA = 0.9
ALPHA = 0.1
NUM_EPISODES = 500000
EPSILON = 1.0
EPSILON_DECAY = 0.9999

def random_action(a, env, eps=0.1):
    p = np.random.random()

    if p < (1 - eps):
        return a
    else:
        return env.action_space.sample()


def play_game(env, epsilon):
    pass

if __name__ == '__main__':
    env = gym.make('FrozenLake-v0')
    env = wrappers.Monitor(env, 'frozenlake-experiment-Q1', force=True)

    Q = np.zeros((env.observation_space.n, env.action_space.n))
    updateCounts = np.zeros((env.observation_space.n, env.action_space.n))

    deltas = []
    averageReturn = 0
    lastHalfAverage = 0
    for episode in xrange(NUM_EPISODES):

        s = env.reset()

        biggest_change = 0
        while True:
            a = np.argmax(Q[s, :])
            a = random_action(a, env, EPSILON)

            old_qsa = Q[s, a]

            s2, r, done, info = env.step(a)
            if (s2==s):
                r-=0.01
            elif(done and r == 0):
                r = -1.0

            #alpha = (ALPHA/(updateCounts[s, a]+1))
            alpha = ALPHA
            updateCounts[s, a] += 1

            a2 = np.argmax(Q[s2, :])
            maxNext = Q[s2, a2]
            Q[s, a] = old_qsa + alpha*(r + GAMMA*maxNext - old_qsa)
            biggest_change = max(biggest_change, np.abs(old_qsa - Q[s, a]))

            updateCounts[s, a] = updateCounts[s, a] + 1

            s=s2
            a=a2

            if(done):
                averageReturn = averageReturn + (r-averageReturn)/(episode+1)
                EPSILON *= EPSILON_DECAY
                if (episode+1)%10000 == 0:
                    print "Episode: ", episode+1
                    print 'Epsilon: ', EPSILON
                    print 'Average Return: ', averageReturn
                break
        deltas.append(biggest_change)

    env.close()