Edit on GitHub

Search Policy

Search policy is a set of types and functions that define the details of the search process. The aim is to make workers focus on the infrastructure of the search, while the policy defines the details of the search algorithm. Several policies may coexist, allowing experimentation with different ideas.

As of 2025-07-25, the search policy defines the following:

Function or struct	Description
MakeNodeKey	Computation of the hash for a position (e.g. whether transpositions end up having the same hash)
GetNumEdgesToFetch	How many edges to consider for the next visit
DistributeVisits	How to distribute visits between child nodes
struct ValueDelta	Structure to use in backpropagation
NodeEventToValueDelta	How to convert NodeEvent to ValueDelta
MergeNodeUpdates	How to combine two ValueDeltas
MoveNodeUpdateToParent	How to forward a ValueDelta to the parent node
UpdateNodeAggregate	How to apply ValueDelta to the node in the repository
MakeEdgeDelta	How to apply ValueDelta to the edge in the repository
(TBD)	Which move to play after the search is finished.

The logic that is currently implemented will be described in the following sections.

The way SearchPolicy is defined is still evolving/subject to change:

• Currently, it is implemented as a struct with static functions, which worker classes incorporate using Policy = SearchPolicy declarations.
• When we need more than one SearchPolicy, one option will be have it as a template parameter of the worker classes, e.g. GatherWorker<SearchPolicy>.
• Another option is to make it a virtual class, and make functions non-static. There may be slight inconvenience with the type definitions though.
• Functions in the SearchPolicy will need access to the engine settings. At first, we’ll just pass the settings as a parameter.
• Later we likely will make it a member variable of the SearchPolicy, meaning the functions will need to be non-static. Note that it doesn’t require functions to be virtual.