How to approach a Machine Learning Evaluation ?

[ekovanda]

Hi,
I'm new to this type of challenge but I understand that you could evaluate positions using some heuristics (.e.g sum of piece value or more elaborate heuristics based on piece positioning etc). This could be used in a recursive minimax.

But how would I conceptually go about a Machine learning evaluation in this challenge? I could imagine training a function to estimate the value of a position using past data and using that to evaluate possible moves. However that would require me to save and call a model which doesn't seem to be allowed by the rules.

Any ideas are appreciated as I would like to try a machine learning approach 😊

[WhiteMouse1]

Find a way to compress the model or find a way to make the resulting model to be small enough to fit into token constrains.

You can try to create smaller models to control smaller parts of your thinking process.
Sorta like insect nervous system instead of mammal's central brain.
One micro-model is fed pawn positions, another is fed available captures, then their outputs are somehow mixed to produce and index of the move to take.

[ekovanda]

@WhiteMouse1, thanks for your response. Typically, I would train and ML model, save it to a file and load it for inference. However, in this case we are not allowed to store files which means I would store the weights directly in an array in the think() function? I'm afraid this will eat up my tokens extremly fast. But I'll give it a shot.

Any further advice is appreciated 😊

[WhiteMouse1]

Yes. Maybe not inside Think function directly, but inside MyBot.cs

Piece-Square Tables are used extensively in this challenge and contestants use compression and decompression algorithms to pack them into least amount of tokens and unpack the values on demand.
This bot has packed PST implemented, for example:
https://github.com/JacquesRW/Chess-Challenge/blob/main/Chess-Challenge/src/My%20Bot/MyBot.cs

I think this type of packing/unpacking data is the direction to take.

[ekovanda]

Thanks for this! I'll give it a go 😊