The PassFilter simply outputs BlackWins and WhiteWins (Range 0 - 1 but not presently clamped).
In principle, this type of feedforward ANN can be used to decide whether a PASS will result in blackwins or whitewins at any stage.
The goal is for the network to learn that passing while losing when valid moves exist is bad, but passing while winning is relatively harmless later in the game.
Now possible to prohibit a Collection of Actions from being included in the List of returned actions (or as a single returned value).
All unit tests now pass with the exception of MonteCarloUCT.
TODO: playerToMove and previousPlayerPassed should be made part of the GameState.
This would remove the superfluous Player parameter from many methods and make it possible
to check for the "I'm ahead and my opponent is offering to end the game" killer move.
Player and Action classes are now singletons (factory pattern) rather than String values.
Implementing more general treesearch code for minimax, alpha-beta, monte carlo using simplified backup logic.
Removed unnecessary distinction between policy and tree search (tree search is a special kind of policy).
Calculation of all valid moves / arbitrary sets of moves is now a seperate class, as it serves a different purpose than a policy.
Introduced regression error in AlphaBeta test.
