Modern Times wrote:Edmund wrote:
If you are planning on running personality tournaments, I am sure we can find a special arrangement.
I had planned to do exactly that with your previous version, but gave up on the idea when it became apparent that any x86 personality is guaranteed to be weaker than the standard x64 exe, so it would have been a waste of time. My plan was to try and move the diehard CM fanatics onto something else as a new challenge for them....
The compile Pawel has just uploaded should fit your needs. It should be almost the same speed as the main compile, but offers a range of personality tweaks. It doesn’t understand the commands that obviously weaken the play, but on the other hand it exposes some of the internal tuning variables that so far have not been tweakable.
I will give a short list of the most relevant parameters for personality creation:
Giving a special weight to certain evaluation features – default: 100
Parameters: Aggression, Caution, Activity, Restraint, Passers, Pawn structure
Code that makes the engine favour blocked positions. With this settings program will want to trade a bishop for a knight in the opening, and then to keep as many pawns as it can. – default: false
Parameter: Blocker
Draw-seeking / avoidance – default: 10
Paramter: Contempt
Adjusting piece values – default: 325, 335, 500, 975
Paramters: Knight, Bishop, Rook, Queen
The value for having the bishop pair – default: 30
Paramter: “Bishop Pair” (without the quotes)
Prefer moving certain pieces over others – default: 0
Parameters: “King play”, “Queen play”, “Rook play”, “Bishop play”, “Knight play”, “Pawn play” (without the quotes)
Tweaking Late Move Reductions
Parameters:
eLMR_mind: at least how many moves have to be tried before starting to reduce: default 4
eLMR_depth: at least how many plies away from the leaves do we have to be: default 5
eLMR_extend_fact & eLMR_extend_static: for deciding on how many plies to reduce. a) reduce x/ eLMR_extend_fact, where x is the number of moves that have been reduced so far b) additionally to a) reduce exactly eLMR_extend_static plies – default: 1024 (= don’t reduce based on x), 4
Tweaking Futility Pruning
Parameters:
eFPRUNE_x1, eFPRUNE_x2: The margin for deciding on whether or not to prune is: x1 * (depth-4)^2 / 100 + x2 – default: 172, 103
Tweaking King Safety
Parameters:
eKS_x1, eKS_x2, eKS_x3: Calculate the King Safety: x1 / ( 1 + 2^( x2 - i / x3 ) ), where i is a variable between 1 and 256 that counts specific features such as sliders attacking the king, rook batteries pointing at the king, etc. The formula describes a S-shaped curve that pays little regard to small number of threats, at a certain point every additional attacker is worth quite a lot and towards the end of the curve when there are already many attackers an additional attacker is not worth so much. x1 is particularly interesting as it gives the maximum value of King Safety – default: 160, 7, 7
Tweaking the Nullmove Reduction
Parameters:
eNULL_x1, eNULL_x2: The reduction depth of nullmove searches equals: x1 / 16 + depth * x2 / 16 – default: 24, 4
Tweaking Lazy Evaluation. First Glass does some basic evaluation of a position like counting material values. If the evaluation is way off the bounds necessary for a cutoff, don’t bother with other more expensive evaluations such as Mobility evaluation.
Parameter: eLAZY_MARGIN – default: 300
