def run(self):
while self.playerIds:
self.game_count += 1
print("Game " + str(self.game_count))
game_session = GameSession(self.playerIds, self.dice_count)
game_session.run()
self.read_players()
def addSession(self, owner, ownerId, gamename):
validAddSession = False
if not self.gameExist(gamename):
newGame = GameSession(owner, ownerId, gamename)
self.sessionArr.append([owner, gamename, newGame])
validAddSession = True
return validAddSession
def runTestPlayRandomGames(numGames, reproSeed):
if reproSeed < 0:
print('Running', numGames, 'random test games...')
for _ in tqdm(range(numGames)):
seed = random.randrange(999999)
random.seed(seed)
np.random.seed(seed)
session = GameSession()
session.randomSeed = seed
while not session.isFinished():
playerId = session.getCurrentPlayer()
state = session.getState(playerId)
actions = session.getAllowedActions(state)
action = random.choice(actions)
session.step(action)
else:
random.seed(reproSeed)
np.random.seed(reproSeed)
session = GameSession()
session.game.debug = True
session.randomSeed = reproSeed
print('DEBUG game')
print(' mods:', [mod.name for mod in session.game.mods])
print(' blue:', session.game.blueDeck)
print(' red:', session.game.redDeck)
while not session.isFinished():
playerId = session.getCurrentPlayer()
state = session.getState(playerId)
actions = session.getAllowedActions(state)
action = random.choice(actions)
print(
'Move, player:%d numActions:%d (board:%d + card:%d) => action:%d (board:%d + card:%d)'
% (playerId, len(actions), sum([state[i] for i in range(9)]),
sum([state[9 + i] for i in range(5)
]), action, action % 9, int(action / 9)))
session.step(action)
def trainModel():
numEpochs = 20
numGamesToTrain = 200
numGamesToEval = 1000
game = GameSession(useModifiers=True)
trainingAgent = AgentDQN(game)
randomAgent = AgentRandom(game)
learningEnv = Environment(game, [trainingAgent, trainingAgent])
evalEnv = Environment(game, [trainingAgent, randomAgent])
# load checkpoint data
#trainingAgent.load('data/model.tmp')
print('Start training, epochs:', numEpochs)
timeStart = time.time()
history = []
for epochIdx in range(numEpochs):
# play a lot and learn and stuff
for _ in tqdm(range(numGamesToTrain),
leave=False,
desc='Training'.ljust(15)):
learningEnv.runTrainingGame()
# training update after exploration step
trainingAgent.train()
# eval with same seed for every iteration
randomAgent.setSeed(0)
score = 0
for _ in tqdm(range(numGamesToEval),
leave=False,
desc='Evaluating'.ljust(15)):
rewards = evalEnv.runEvalGame()
score += rewards[0]
# collect info about training to see if it's actually working (lol, who am i kidding, it just fails constantly T.T)
stepDetails = trainingAgent.getTrainingDetails()
stepDetails['score'] = score
history.append(stepDetails)
print('[%d] %s' % (epochIdx, stepDetails))
timeElapsed = time.time() - timeStart
print('Done. Total time:%s, epoch avg:%.0fs' % (time.strftime(
"%H:%M:%Ss", time.gmtime(timeElapsed)), timeElapsed / numEpochs))
trainingAgent.save('data/model.tmp')
trainingAgent.generateModelCode('data/model.cs')
return history
def runTestAgentEval(numGames, agentName):
print('Running', numGames, 'eval games for agent', agentName)
game = GameSession()
evalAgent = AgentRandom(game)
if agentName == 'dqn':
evalAgent = AgentDQN(game)
env = Environment(game, [evalAgent, AgentRandom(game)])
score = 0
for _ in tqdm(range(numGames)):
rewards = env.runEvalGame()
score += rewards[0]
""" Example of use of GameSession and Engine to make Stockfish play itself. """ from gameSession import GameSession from engine import Engine stockfishpath = './Stockfish/src/stockfish' eng1 = Engine(stockfishpath) eng2 = Engine(stockfishpath) GameSession(eng1, eng2).play() GameSession(eng1, eng2, time=(300e3, 5e3)).play()