exploration

Type Members

1. case class BootstrapModel[U, N, A, B <: U](modelId: ModelIdentity, models: IndexedSeq[Submodel[Int, A, U]], salt: GenAggFunc[A, Long], classLabels: IndexedSeq[N], auditor: Auditor[U, N, B]) extends SubmodelBase[U, N, A, B] with Product with Serializable

   

   A model for performing bootstrap style exploration.

   A model for performing bootstrap style exploration. This makes use of a number of policies. The algorithm chooses one policy and then uses the other to calculate the appropriate probability of choosing that action. Note that the models MUST return a value between 1 and the number of actions, and if not an exception will be thrown.

   A

   model input type

   B

   model output type

   modelId

   a model identifier

   models

   a set of models that generate Int's. These models MUST be deterministic for the probability to be correct. Each model must return a value in the range 1 to classLabels.size (inclusive).

   salt

   a function that generates a salt for the randomization layer. This salt allows the random choice of which policy to follow to be repeatable.

   classLabels

   a list of class labels to output for the final type. Also note that the size of this controls the number of actions. If the submodel returns a score < 1 or > classLabels.size (note the 1 offset) then a RuntimeException will be thrown.

2. case class EpsilonGreedyModel[U, N, -A, B <: U](modelId: ModelIdentity, defaultPolicy: Submodel[Int, A, U], epsilon: Float, salt: GenAggFunc[A, Long], classLabels: IndexedSeq[N], auditor: Auditor[U, N, B]) extends SubmodelBase[U, N, A, B] with Product with Serializable

   

   A model which does epsilon greedy style exploration.

   A model which does epsilon greedy style exploration. This will choose a random action with probability epsilon or an action from the defaultPolicy with probability 1 - epsilon. Note that the default policy MUST return a value between 1 and the number of actions, and if not an exception will be thrown.

   A

   model input type

   B

   model output type

   modelId

   a model identifier

   defaultPolicy

   the model to use for exploitation. This MUST be deterministic for the probability to be correct. The model must return a value in the range 1 to classLabels.size (inclusive).

   epsilon

   the exploration/exploitation tradeoff parameter. epsilon must be in the interval [0, 1]. 0 indicates never select an action randomly. 1 indicates always select an action randomly.

   salt

   a function that generates a salt for the randomization layer. This salt allows the random choice of which policy to follow to be repeatable.

   classLabels

   a list of class labels to output for the final type. Also note that the size of this controls the number of actions. If the submodel returns a score < 1 or > classLabels.size (note the 1 offset) then a RuntimeException will be thrown.