The Chef's Hat Competition Agens provides four implementation agents for the Chef's Hat game.
The instalation can be done via pip:
pip install -r requirements.txtor just run setup.py
python setup.py installor
pip install .There are four different implementations of Agents, all of them is based on the Karma Chameleon Club
We have used all the available dataset in the Chef's Hat Player Club envinroment to train the agents:
| Types | Types | Types | Types | Types |
|---|---|---|---|---|
| lil_abcd_ | lilAbsol | lilAle | lilAna | lilArkady |
| lilAuar | lilBlio1 | lilBlio2 | lilChu | lilDa48 |
| lilDana | lilDJ | lilDomi948 | lilEle | lilFael |
| lilGeo | lilIlzy | lilJba | lilLeandro | lilLena |
| lilLordelo | lilMars | lilNathalia | lilNik | lilNilay |
| lilRamsey | lilRaos | lilThecube | lilThuran | lilTisantana |
| lilToran | lilWinne | lilYves | lilYves2 |
We just concatenate all available dataset in only one new dataset.
For the first agent, we decided to use a custom loss function based on the Hub Loss and the following getRewarded function
def getReward(self, info, stateBefore, stateAfter):
thisPlayer = info["thisPlayerPosition"]
matchFinished = info["thisPlayerFinished"]
maxScore = numpy.argmax(info["score"], axis=0)
beforeScore = self.beforeScore
beforeInfo = self.beforeInfo
self.beforeInfo = info
self.beforeScore = info["score"][0]
if matchFinished:
if thisPlayer == 0:
return 1
else:
if beforeScore is not None and info["score"][0] > beforeScore:
if maxScore == 0:
return 1
else:
return 0.001
return -0.001
state = numpy.concatenate((stateBefore[0:11], stateBefore[11:28]))
reward = -0.001
if beforeInfo is not None:
lastActionPlayers = beforeInfo["lastActionPlayers"]
actualActionPlayers = info["lastActionPlayers"]
count = 0
if len(lastActionPlayers[0]) > 0 and lastActionPlayers[0][0] == 'DISCARD':
for x, i in enumerate(actualActionPlayers):
if x == 0:
continue
for val in actualActionPlayers[x]:
if len(actualActionPlayers[0]) > 0 and actualActionPlayers[0][0] == 'DISCARD' and (val == 'PASS' or val == ''):
reward += 0.05
count += 1
if len(actualActionPlayers[0]) > 0 and actualActionPlayers[0][0] != 'PASS' and count == 3:
return 1
if beforeInfo is not None:
if len(lastActionPlayers[0]) > 0 and lastActionPlayers[0][0] == 'DISCARD' and len(
actualActionPlayers[0]) > 0 and actualActionPlayers[0][0] == 'DISCARD':
reward += 0.05
if reward > 1:
return 1
return rewardEach agent can be instantiated using:
agent = AMYG4.AMYG4(name="AMYG4", continueTraining=False, initialEpsilon=1, verbose=True) # training agentFor the our second agent implementation we just use the following getRewarded function:
def getReward(self, info, stateBefore, stateAfter):
thisPlayer = info["thisPlayerPosition"]
matchFinished = info["thisPlayerFinished"]
if matchFinished:
if thisPlayer == 0:
return 1
else:
return -0.001
state = numpy.concatenate((stateBefore[0:11], stateBefore[11:28]))
rewardShape = numpy.concatenate([state, info["action"]])
rewardShape = numpy.expand_dims(numpy.array(rewardShape), 0)
reward = self.rewardNetwork.predict([rewardShape])[0][0]
if self.beforeInfo is not None:
lastActionPlayers = self.beforeInfo["lastActionPlayers"]
actualActionPlayers = info["lastActionPlayers"]
count = 0
if len(lastActionPlayers[0]) > 0 and lastActionPlayers[0][0] == 'DISCARD':
for x, i in enumerate(actualActionPlayers):
if x == 0:
continue
for val in actualActionPlayers[x]:
if len(actualActionPlayers[0]) > 0 and actualActionPlayers[0][0] == 'DISCARD' and val == 'PASS':
reward += 0.5
count += 1
if len(lastActionPlayers[0]) > 0 and lastActionPlayers[0][0] != 'PASS' and count == 3:
return 1
self.beforeInfo = info
if reward > 1:
return 1
return rewardEach agent can be instantiated using:
agent = AINSA.AINSA(name="AINSA", continueTraining=False, initialEpsilon=1, verbose=True) # training agentThird implementation is only based on the the following getRewarded function:
def getReward(self, info, stateBefore, stateAfter):
thisPlayer = info["thisPlayerPosition"]
matchFinished = info["thisPlayerFinished"]
if matchFinished:
if thisPlayer == 0:
return 1
else:
return -0.001
state = numpy.concatenate((stateBefore[0:11], stateBefore[11:28]))
rewardShape = numpy.concatenate([state, info["action"]])
rewardShape = numpy.expand_dims(numpy.array(rewardShape), 0)
reward = self.rewardNetwork.predict([rewardShape])[0][0]
return rewardEach agent can be instantiated using:
agent = AIACIMP.AIACIMP(name="AIACIMP", continueTraining=False, initialEpsilon=1, verbose=True) # training agentFinally, the last is rewarded only if win:
def getReward(self, info, stateBefore, stateAfter):
thisPlayer = info["thisPlayerPosition"]
matchFinished = info["thisPlayerFinished"]
if matchFinished:
if thisPlayer == 0:
return 1
return -0.001Each agent can be instantiated using:
agent = ALLIN.ALLIN(name="ALLIN", continueTraining=False, initialEpsilon=1, verbose=True) # training agent| Parameter | Description |
|---|---|
| name | name of the agent, must be unique in a game |
| continueTraining | if the agent will learn during the game. Must have a demonstration if continue learning is True. |
| demonstrations | must be a npy with a list of ([state, action, possibleActions]). The default value is the dataset foleder inside the agent folder |
| initialEpsilon | when learning, initial exploration value |
| saveFolder | folder that the agent will be saved in. The default value is the same folder of the agent |
| verbose | verbose when learning |
Yves Galvão - ymg@ecomp.poli.br