def test_cur_state_when_opponent_should_move_to_start(self):
game = TicTacToe([['X', 'O', 'O'], ['X', 'O', ' '], [' ', 'X', 'X']])
ab = AlphaBeta()
mdp = FixedGameMDP(game, ab, 1)
env = Environment(mdp)
expected = TicTacToe([['X', 'O', 'O'], ['X', 'O', ' '],
['O', 'X', 'X']])
self.assertEqual(env.cur_state(), expected)
def test_do_action(self):
# X - O
# - - X
# - - O
game = TicTacToe().make_moves([1, 3, 6, 9])
mdp = FixedGameMDP(game.copy(), AlphaBeta(), 1)
env = Environment(mdp)
env.do_action(7)
expected = TicTacToe().make_moves([1, 3, 6, 9, 7, 4])
self.assertEqual(env.cur_state(), expected)
def on_episode_begin(self, episode, qfunction):
mdp = FixedGameMDP(get_random_game(), RandPlayer(random_state=seed), 1)
env = Environment(mdp)
qlearning.env = env
egreedy.action_space = env.actions
qlearning.policy.provider = env.actions
if episode % 50 == 0:
print('Episode {}'.format(episode))
from capstone.game.games import Connect4 as C4
from capstone.game.players import RandPlayer
from capstone.rl import Environment, GameMDP, FixedGameMDP
from capstone.rl.learners import ApproximateQLearning as ApproxQLearning
from capstone.rl.policies import EGreedy, RandomPolicy
from capstone.rl.utils import EpisodicWLDPlotter, Callback, LinearAnnealing
from capstone.rl.value_functions.c4deepnetwork import Connect4DeepNetwork
import numpy as np
import random
seed = 383
random.seed(seed)
np.random.seed(seed)
mdp = FixedGameMDP(get_random_game(), RandPlayer(random_state=seed), 1)
env = Environment(mdp)
c4dn = Connect4DeepNetwork()
egreedy = EGreedy(action_space=env.actions,
qfunction=c4dn,
epsilon=1.0,
selfplay=False,
random_state=seed)
qlearning = ApproxQLearning(env=env,
qfunction=c4dn,
policy=egreedy,
discount_factor=0.99,
selfplay=False,
experience_replay=True,
replay_memory_size=20000,
batch_size=32)
from capstone.game.games import TicTacToe
from capstone.game.players import RandPlayer
from capstone.rl import GameMDP, FixedGameMDP, Environment
from capstone.rl.learners import ApproximateQLearning
from capstone.rl.policies import RandomPolicy
from capstone.rl.utils import EpisodicWLDPlotter, QValuesPlotter
from capstone.rl.value_functions import MLP
seed = 23
game = TicTacToe()
env = Environment(FixedGameMDP(game, RandPlayer(), 1))
mlp = MLP()
qlearning = ApproximateQLearning(
env=env,
policy=RandomPolicy(env.actions, random_state=seed),
qfunction=mlp,
discount_factor=1.0,
n_episodes=50000
)
qlearning.train(
callbacks=[
EpisodicWLDPlotter(
game=game,
opp_player=RandPlayer(random_state=seed),
n_matches=1000,
period=5000,
# filepath='../mlnd-capstone-report/figures/tic_ql_tab_full_selfplay_wld_plot.pdf'
filepath='figures/test88.pdf'
)
]
)
def test_cur_state(self):
game = TicTacToe()
mdp = FixedGameMDP(game, AlphaBeta(), 1)
env = Environment(mdp)
self.assertEqual(env.cur_state(), mdp.start_state())
self.assertEqual(env.cur_state(), game)