MACS Parameters

MACS is a flexible system thanks to a set of parameters that can be modified without recompiling the system. These parameters are stored in a file whose name must be given to the constructor of MACS so that they are initialized correctly. The parameters can be written in any order. Some of them are deprecated. In order to set the value of a parameter, write in the file a line such as:

parameterName = value

The list of current parameters is in the following table. The sample values given correspond to a set-up where active exploration is used and MACS performs 10 random planning steps per time step. You can generally use one of the MACS parameters file "as such".

The active exploration parameters are stored in a list. The user must first give the number of criteria, and then, for each criterion, give its name (among "info","rehearsal","stateSearch" and "external", so far), followed by its rank number, and some parameters going with these criteria. The order is important: criterion number one will have priority over criterion number two, and so on.

parameter namesample valuemeaning
systemNameMACSCould be YACS or any other...
General learning parameters
initialReliability0.5The value of classifier reliability when created
learningRate0.5The famous beta in Widrow-Hoff
eGreedy0.0epsilon value for exploration
nbPlanningSteps10number of value iteration steps on the model as in Dyna
plannerTypeRandomPlanner or SystematicPlannerif random, nbPlanningSteps chosen randomly ; if systematic, nbPlanningSteps iterations on all known states
classifierMemorySize5number of good and bad markers recorded
nbEvaluationsToSpecialize5
nbFailuresToRemove5
nbSuccessToProbe5
Active exploration parametersUsed for active exploration. See above
nbCriteria3
criteriaNames0infothe information reward
criteriaDiscountFactor00.5
criteriaDefaultValue01.0
criteriaIncrLearningRate00.1
criteriaDecrLearningRate01.0
criteriaNames1externalthe external reward
criteriaDiscountFactor10.5
criteriaDefaultValue10.0
criteriaIncrLearningRate11.0
criteriaDecrLearningRate11.0
criteriaNames2rehearsalthe rehearsal reward
criteriaDiscountFactor20.5
criteriaDefaultValue20.0
criteriaIncrLearningRate20.1
criteriaDecrLearningRate21.0
General switches
conditionCoveringon
obsoleteConditionSelectionon??? still used ?
anticipationLearningon
mutspecDiscoveringon
generalizationon
Trace parameters
traceConditionCoveringonAdds a message each time condition covering creates a classifier
traceObsoleteConditionSelectionon??? still used ?
traceAnticipationCoveringon??? still used ?
traceAnticipationSelectionon??? still used ?
traceMutspecCreationonAdds a message each time mutspec creates a classifier
traceMutspecRemovalonAdds a message each time mutspec removes a classifier
traceMutgenCreationonAdds a message each time generalization creates a classifier
traceMutgenRemovalonAdds a message each time generalization removes a classifier
traceStatisticsoff? error bars ?
traceGeneralInfosonnb classifiers, nb good, % knowledge...
traceInputsonshows the input at each time step
traceOutputonshows the action chosen at each time step
traceClassifierListFrequency1000shows the complete classifier list at each interval
traceSituationValuesFrequency1000shows the values of all known states at each interval
traceTimeonshows the number of time steps