|
| 1 | +import gym |
| 2 | +import itertools |
| 3 | +import matplotlib |
| 4 | +import matplotlib.style |
| 5 | +import numpy as np |
| 6 | +import pandas as pd |
| 7 | +import sys |
| 8 | + |
| 9 | + |
| 10 | +from collections import defaultdict |
| 11 | +from windy_gridworld import WindyGridworldEnv |
| 12 | +import plotting |
| 13 | + |
| 14 | +matplotlib.style.use('ggplot') |
| 15 | + |
| 16 | +# transition probability {0,1} |
| 17 | +# Reward {-1,0,1} |
| 18 | +# States {S0,S1,S2,S3,S4,S5,..........SN} |
| 19 | +# Acation {left, Right,Down, Up} |
| 20 | + |
| 21 | + |
| 22 | +# Question |
| 23 | +""" |
| 24 | +1. what are the Good State (Value Funcation) |
| 25 | +2. What are Good State and Action pair(Q-Value Funcation) |
| 26 | + |
| 27 | +Take these Files from Github |
| 28 | +https://github.com/reddyprasade/Machine-Learning-with-Scikit-Learn-Python-3.x/tree/master/Reinforcement%20Learning/Q-Learning |
| 29 | + |
| 30 | +1. windy_gridworld.py |
| 31 | +2. plotting.py |
| 32 | +""" |
| 33 | + |
| 34 | +env = WindyGridworldEnv() |
| 35 | + |
| 36 | + |
| 37 | +## Make the $\epsilon$-greedy policy. |
| 38 | + |
| 39 | + |
| 40 | +def createEpsilonGreedyPolicy(Q, epsilon, num_actions): |
| 41 | + """ |
| 42 | + Creates an epsilon-greedy policy based |
| 43 | + on a given Q-function and epsilon. |
| 44 | + |
| 45 | + Returns a function that takes the state |
| 46 | + as an input and returns the probabilities |
| 47 | + for each action in the form of a numpy array |
| 48 | + of length of the action space(set of possible actions). |
| 49 | + """ |
| 50 | + def policyFunction(state): |
| 51 | + |
| 52 | + Action_probabilities = np.ones(num_actions, |
| 53 | + dtype = float) * epsilon / num_actions |
| 54 | + |
| 55 | + best_action = np.argmax(Q[state]) # for Which State which action is best |
| 56 | + Action_probabilities[best_action] += (1.0 - epsilon) |
| 57 | + return Action_probabilities |
| 58 | + |
| 59 | + return policyFunction |
| 60 | + |
| 61 | + |
| 62 | + |
| 63 | + |
| 64 | + |
| 65 | + |
| 66 | +# Build Q-Learning Model. |
| 67 | + |
| 68 | +def qLearning(env, num_episodes, discount_factor = 1.0, |
| 69 | + alpha = 0.6, epsilon = 0.1): |
| 70 | + """ |
| 71 | + Q-Learning algorithm: Off-policy TD control. |
| 72 | + Finds the optimal greedy policy while improving |
| 73 | + following an epsilon-greedy policy""" |
| 74 | + |
| 75 | + # Action value function |
| 76 | + # A nested dictionary that maps |
| 77 | + # state -> (action -> action-value). |
| 78 | + Q = defaultdict(lambda: np.zeros(env.action_space.n)) |
| 79 | + |
| 80 | + # Keeps track of useful statistics |
| 81 | + stats = plotting.EpisodeStats( |
| 82 | + episode_lengths = np.zeros(num_episodes), |
| 83 | + episode_rewards = np.zeros(num_episodes)) |
| 84 | + |
| 85 | + # Create an epsilon greedy policy function |
| 86 | + # appropriately for environment action space |
| 87 | + policy = createEpsilonGreedyPolicy(Q, epsilon, env.action_space.n) |
| 88 | + |
| 89 | + # For every episode |
| 90 | + for ith_episode in range(num_episodes): |
| 91 | + |
| 92 | + # Reset the environment and pick the first action |
| 93 | + state = env.reset() |
| 94 | + |
| 95 | + for t in itertools.count(): |
| 96 | + |
| 97 | + # get probabilities of all actions from current state |
| 98 | + action_probabilities = policy(state) |
| 99 | + |
| 100 | + # choose action according to |
| 101 | + # the probability distribution |
| 102 | + action = np.random.choice(np.arange( |
| 103 | + len(action_probabilities)), |
| 104 | + p = action_probabilities) |
| 105 | + |
| 106 | + # take action and get reward, transit to next state |
| 107 | + next_state, reward, done, _ = env.step(action) |
| 108 | + |
| 109 | + # Update statistics |
| 110 | + stats.episode_rewards[ith_episode] += reward |
| 111 | + stats.episode_lengths[ith_episode] = t |
| 112 | + |
| 113 | + # TD Update |
| 114 | + best_next_action = np.argmax(Q[next_state]) |
| 115 | + td_target = reward + discount_factor * Q[next_state][best_next_action] |
| 116 | + td_delta = td_target - Q[state][action] |
| 117 | + Q[state][action] += alpha * td_delta |
| 118 | + |
| 119 | + # done is True if episode terminated |
| 120 | + if done: |
| 121 | + break |
| 122 | + |
| 123 | + state = next_state |
| 124 | + |
| 125 | + return Q, stats |
| 126 | + |
| 127 | + |
| 128 | +# Now i want to train the model |
| 129 | + |
| 130 | +Q,stats = qLearning(env,5) |
| 131 | + |
| 132 | +# Plot important statistics. |
| 133 | +plotting.plot_episode_stats(stats) |
0 commit comments