In [ ]:
%matplotlib inline

In [ ]:
from windy import WindyGridWorld
import numpy as np
from collections import deque
import pickle
import matplotlib.pyplot as plt

Necessary functions


In [ ]:
def epsilon_greedy_policy(q_vector, epsilon):
    is_greedy_action = False if np.random.uniform() <= epsilon else True
    if is_greedy_action:
        #Random choice, if there are at least 2 Q-s with the same values
        # TODO: Check, if randomizing the q_vector in lookup_action_value() around 0, what would happen
        max_q_args = np.argwhere(q_vector == np.amax(q_vector))
        if len(max_q_args) > 1:
            action = np.random.choice(max_q_args.ravel(), 1)[0] + 1
        else:
            action = np.argmax(q_vector) + 1
    else:
        action = np.random.randint(1, 5)
    return action

In [ ]:
def td_error(c_a_v, n_a_v, r, g=0.95):
    return r + g * n_a_v - c_a_v

In [ ]:
def action_value(feature_vector, weights):
    return feature_vector @ weights.T

In [ ]:
def get_action_value_vector(state, weights):
    val = []
    for act in range(1, 5):
        val.append(get_action_value(state, weights, act))
    return val

In [ ]:
def calc_distance(state_vector, winner_tile):
    return np.linalg.norm(np.asarray(state_vector)-np.asarray(winner_tile))

In [ ]:
def feature_vector(state_vector, action):
    return np.hstack((1 / (np.asarray(state_vector) + (np.asarray([-1, 0]), np.asarray([0, 1]), np.asarray([1, 0]), np.asarray([0, -1]))[action - 1] + np.array([1, 1])), 1 / np.asarray(state_vector)))

In [ ]:
def get_action_value(s, w, a):
    return action_value(feature_vector(s, a), w)

In [ ]:
def td_lambda_update(curr_state, next_state, action, w, e_trace, reward, epsilon, _lambda=0.9, gamma=0.95, alpha=0.15):
    # Get current action-value (belonging to current state and action)
    current_action_value = get_action_value(curr_state, w, action)
    # Get the next state action-value vector
    next_action_value_vector = get_action_value_vector(next_state, w)
    # Get the next action according to the next_action_value_vector and epsilon(Using epsilon-greedy policy with current epsilon)
    next_action = epsilon_greedy_policy(next_action_value_vector, epsilon)
    # Get next action-value from the next_action_value_vector according our action
    next_action_value = next_action_value_vector[next_action - 1]
    
    ## UPDATE THE WEIGHTS ##
    
    # Calculate the TD-error
    delta = td_error(current_action_value, next_action_value, reward, gamma)
    # Update eligibility trace
    new_e_trace = gamma * _lambda * e_trace + feature_vector(curr_state, action)
    # Weights delta
    delta_w = alpha * delta * new_e_trace
    
    return next_action, delta_w, new_e_trace

In [ ]:
def try_table(w):
    game = WindyGridWorld(GRID_SIZE, WINNER_TILE, WINDY_ARRAY, START_TILE)
    current_pos = game.current_pos()
    is_ended = False
    agent_positions = deque()
    states_reward_list = []
    actions_deque = deque()
    epsilon = 0.01
    
    while not is_ended:
        
        # Append current agent position
        agent_positions.append(current_pos)

        # Lookup action value belonging to current state
        action_value_vector = get_action_value_vector(current_pos, w)

        # Choose action based on current action value vector and epsilon
        action = epsilon_greedy_policy(action_value_vector, epsilon)
        # Append action
        actions_deque.append(action)

        # Get next state, reward
        current_pos, reward, is_ended = game.step(action)
    
    return agent_positions, actions_deque, current_pos

In [ ]:
def show_moves(visited_states, grid_size):
    arr = np.asarray(visited_states).T
    range_x = (0.5, grid_size[1] + 0.5)
    range_y = (0.5, grid_size[0] + 0.5)
    ax = plt.gca()
    ax.scatter(arr[1], arr[0])
    ax.quiver(arr[1,:-1],arr[0,:-1],arr[1,1:]-arr[1,:-1],arr[0,1:]-arr[0,:-1], scale_units='xy', angles='xy', scale=1)
    ax.set_xticks(np.arange(*range_x), minor=True)
    ax.set_yticks(np.arange(*range_y), minor=True)
    ax.set_xlim(*range_x)
    ax.set_ylim(*range_y)
    ax.set_xlabel("Valami")
    ax.invert_yaxis()
    ax.get_xaxis().set_tick_params(labeltop="on", labelbottom="off")
    ax.set_aspect("equal")
    plt.grid(which="minor")
    plt.show()

Game parameters


In [ ]:
GRID_SIZE = (6, 6)
WINNER_TILE = (3, 6)
# WINDY_ARRAY = (0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 2, 2, 0)
WINDY_ARRAY = np.zeros(6)
START_TILE = (1, 1)

Main learning loop


In [ ]:
feature_vector_size = 4
W = np.zeros(feature_vector_size)

epsilon = 1
episode_number = 0

average = 0
alpha = 0.15
episodes_number=200
#Külső ciklus
for curr_episode in range(episodes_number):
    print("-", end="")
    episode_number += 1
    alpha = alpha - alpha/episodes_number* curr_episode
    if episode_number % 1000 == 0:
        print(str(episode_number) + ". játék")
        print("Lépések száma: " + str(average / 1000))
        average = 0

    epsilon = 1 / (episode_number**(1/3))
    # Containers
    agent_positions = deque()
    states_reward_list = []
    actions_deque = deque()
    # End bool
    is_ended = False
    
    # Reset eligibility traces to zero
    ELIGIBILITY_TRACE = np.zeros(feature_vector_size)
    
    # New game
    game = WindyGridWorld(GRID_SIZE, WINNER_TILE, WINDY_ARRAY, START_TILE)
    current_pos = game.current_pos()
    # TODO: check if current_pos type (tuple) is a problem or not
    # Get action value belonging to current state (through feature vectors)
    action_value_vector = get_action_value_vector(current_pos, W)
    # Choose action based on current action value vector and epsilon
    action = epsilon_greedy_policy(action_value_vector, epsilon)
    
    #Belső ciklus
    while not is_ended:
    #for _ in range(15):
        
        # Append current agent position
        agent_positions.append(current_pos)

        # Append action
        actions_deque.append(action)

        # Get next state, reward
        current_pos, reward, is_ended = game.step(action)

        # Append reward
        states_reward_list.append(reward)
        
        # TD lambda update
        action, delta_W, ELIGIBILITY_TRACE = td_lambda_update(agent_positions[-1], current_pos, action, W, ELIGIBILITY_TRACE, reward, epsilon)
        W += delta_W
    
    # TODO: Something is wrong with this, dunno what yet
    vertical_distance = WINNER_TILE[0] - agent_positions[0][0]
    horizontal_distance = WINNER_TILE[1] - agent_positions[0][1]
    optimal = max((horizontal_distance, vertical_distance))
    average += len(actions_deque) / optimal

In [ ]:
get_action_value_vector((2, 6), W)

Visualization


In [ ]:
a, b, c = try_table(W)
a.append(c)
show_moves(a, GRID_SIZE)

In [ ]:
count = 0
for _ in range(4000):
    a,b,c = try_table()
    count += len(b) / 19
count / 4000