def f(self, player):
if not isinstance(player, CapturePlayer):
player = ModuleDecidingPlayer(player,
self.env,
greedySelection=True)
player.color = CaptureGame.WHITE
self.opponent.color = CaptureGame.BLACK
self.reset()
current = 0
games = 0
while games < self.maxGames and not self.stable(current):
games += 1
self.env.reset()
self.env.giveHandicap(current, self.opponent.color)
self.env.playToTheEnd(self.opponent, player)
win = self.env.winner == player.color
self.addResult(current, win, self.env.movesDone)
if self.goUp(current) and current < self.maxHandicaps:
current += 1
elif self.goDown(current) and current > 1:
current -= 1
high = self.bestHandicap()
# the scale goes from -1 to (the highest handicap + 1)
if not self.fluctuating():
return high + self.winProp(high)
else:
return (high -
0.5) + (self.winProp(high) + self.winProp(high - 1)) / 2.
def __call__(self, p1, p2):
self.temp = self.minTemperature
if self.useNetworks:
p1 = ModuleDecidingPlayer(p1, self.task.env, temperature=self.temp)
p2 = ModuleDecidingPlayer(p2, self.task.env, temperature=self.temp)
else:
assert isinstance(p1, CapturePlayer)
assert isinstance(p2, CapturePlayer)
p1.game = self.task.env
p2.game = self.task.env
p1.color = CaptureGame.BLACK
p2.color = -p1.color
self.player = p1
self.opponent = p2
# the games with increasing temperatures and lower coefficients
coeffSum = 0.
score = 0.
np = int(self.cases * (1 - self.presetGamesProportion))
for i in range(self.maxGames):
coeff = 1 / (10 * self.temp + 1)
preset = None
if self.cases > 1:
if i % self.cases >= np:
preset = self.sPos[(i - np) % self.cases]
elif i < self.cases:
# greedy, no need to repeat, just increase the coefficient
if i == 0:
coeff *= np
else:
continue
res = self._oneGame(preset)
score += coeff * res
coeffSum += coeff
if self.cases == 1 or (i % self.cases == 0 and i > 0):
self._globalWarming()
return score / coeffSum
def f(self, x):
""" If a module is given, wrap it into a ModuleDecidingAgent before evaluating it.
Also, if applicable, average the result over multiple games. """
if isinstance(x, Module):
agent = ModuleDecidingPlayer(x, self.env, greedySelection = True)
elif isinstance(x, CapturePlayer):
agent = x
else:
raise NotImplementedError('Missing implementation for '+x.__class__.__name__+' evaluation')
res = 0
agent.game = self.env
self.opponent.game = self.env
for _ in range(self.averageOverGames):
agent.color = -self.opponent.color
x = EpisodicTask.f(self, agent)
res += x
return res / float(self.averageOverGames)
__author__ = 'Tom Schaul, [email protected]' from pybrain.rl.environments.twoplayergames import CaptureGame from pybrain.rl.environments.twoplayergames.capturegameplayers import RandomCapturePlayer, KillingPlayer, ModuleDecidingPlayer from pybrain.rl.environments.twoplayergames.capturegameplayers.clientwrapper import ClientCapturePlayer from pybrain.rl.experiments.tournament import Tournament from pybrain.tools.shortcuts import buildNetwork from pybrain import SigmoidLayer game = CaptureGame(5) randAgent = RandomCapturePlayer(game, name = 'rand') killAgent = KillingPlayer(game, name = 'kill') # the network's outputs are probabilities of choosing the action, thus a sigmoid output layer net = buildNetwork(game.outdim, game.indim, outclass = SigmoidLayer) netAgent = ModuleDecidingPlayer(net, game, name = 'net') # same network, but greedy decisions: netAgentGreedy = ModuleDecidingPlayer(net, game, name = 'greedy', greedySelection = True) agents = [randAgent, killAgent, netAgent, netAgentGreedy] try: javaAgent = ClientCapturePlayer(game, name = 'java') agents.append(javaAgent) except: print('No Java server available.') print() print('Starting tournament...') tourn = Tournament(game, agents)