This library explores the possibility of using Monte Carlo Tree Search for arbitrary combinatorial search problems. Some search problems have interdependent costs, which is difficult to model using standard ILP approaches. The probabilistic nature of the MCTS tree makes it suitable for these sorts of problems.
DABTree under research to solve route guidance problems with a social equilibrium objective. It extends MCTS
DABTree depends on the following Scala libraries:
See releases for .jar files.
Consider the example in com.github.robfitzgerald.dabtree.local.example.CombSearchTests. The problem is to find a vector of values, each in the range [0,1].
To do this in a local context (storing tree operations in a List and performing samples synchronously), a LocalSyncSearch can be called.
The user must define/provide the following functions and parameters:
- simulate: adds (arbitrarily) to a partial solution state to make it a complete solution state
- evaluate: determines the cost of a solution state
- objective: Minimize, or Maximize
- generateChildren: generates possible child states by adding Actions to a partial solution state
- rankingPolicy: ranks a payload based on its state, reward, stats, or other details. a few example generic Ranking functions are provided.
- allowChildPromotion: a function that is true if promotion is allowed for the provided child.
- activatedPayloadLimit: number of payloads that can be sampling each iteration, based on rank
- totalPayloadCapacity: the number beyond which payloads will be permanently cancelled based on rank
- startFrontier: the position to start the search from; for example, if the state type is a
List[Double]and Action aDouble, starting from an empty tree is providingList((List(), Option.empty[Double])).