This file describes parameters for the decision tree computation with prism_dectree.jar

To enable decision tree learning after a heuristic computation, add the paramter -dectree[,option]*
options are:

train=integer:
	Number of delta* simulations for training the decision tree. 
	default: 1000

verify=integer:
	Number of simulations done to test the decision tree strategy. 
	default: 1000

search:
	Perform a binary search for best M parameter starting from minimum number of good and bad instances learnt.

search=integer:
	Perform a binary search for best M parameter starting from twice the input (which should be a best guess).

onlyTarget:
	When specified, only successful simulations are learned (corresponds to R in the paper).
	Otherwise all simulations are learned (corresponds to A in the paper).

multiLearn:
	When specified, the amount of times a state is visited during simulations 
	is taken into account when training (corresponds to P in the paper).
	Otherwise every visited state is counted exactly once (corresponds to E in the paper).

bdd:	
	When specified, one evaluation run is attached 
	
stratasDTSim:
	When specified, a PRISM strategy for the current problem is generated and trained as a decision tree
	by performing simulations similar to the case of BRTDP.
	Using simulations has the benefit of getting being able to use the importance measure.
	
stratasDTFull:
	When specified, a PRISM strategy for the current problem is generated and trained as a decision tree
	by iterating over the state space and learning the state/action pair specified by the strategy.
	Quantitative information about the importance of the state can not be gathered this way.
	
learn=0 | 1 | 2:
	Specifies the learning mode.
	0 learns state/action pairs according to the importance of the state (corresponds to I in the paper).
	1 learns so that overall equally many good and bad state/action pairs are learned.
	2 learns all state/action pairs exactly once (corresponds to O in the paper).
	default: 0

productEval:
	Evaluates decision trees by model checking the markov chain that is the product of the strategy and the explicit model.
	Note that this approach is only feasible if the explicit model can be constructed.

minIt=integer, confIt=integer:
	In case no search is performed and minIt and/or confIt is set, a number of M and C values are tried.
	The first M value is 2, the last value is the minimum number of good and bad instances learnt.
	The intermediate M values are scaled exponentially.
	The first C value is 0.0001, the last value is 0.5.
	The intermediate C values are scaled linearly.

min=integer,conf=double:
	Sets the default min and conf parameters.
	min is only used in case no search or minIt was specified.
	conf is used also for the binary search
	default: 2 / 0.0001

export=string:
	File name in which the experimental evaluation should be written as a csv file. 
	If no export option is given, then the data is written to stdout.