Multi-move games are really far harder for the AI than single-move games. Because of its generality and the fact it is written in JavaScript, the Diagram's AI is quite slow (it searches only 10k positions per sec on a PC, while programs like Fairy-Max or ChessV typically do 500k/sec). So it cannot look very far ahead; adjusting the depth to 4 ply already makes it unbearably slow for large variants. But in a 3-move game thinking 4 ply ahead does not even bring the complete opponent's turn within the horizon. One could save a factor by not searching move sequences that are merely permutations of an already searched line (which for a 3-move game would save a factor 3! = 6 per turn), which would be automatic if the AI used a transposition table. (But the current one doesn't.) And moves cannot always be made in a prescribed order, as they could block each other even when they must be made by different pieces.
Large games often 'factorize' in a number of independent battles, but it is difficult for an AI to exploit that (but easy for a human): even in a single-move game the AI will try to interleave the moves of the varios battles in combinatorially many ways.
I could do a two-move game like Duck Chess (with dedicated code) only because it is not really a 2-move game, but a Refusal variant.
So I guess for now an AI for multi-move games are not really feasible. Even piece drops are too difficult at the current stage, and solving this has a much higher priority.
Multi-move games are really far harder for the AI than single-move games. Because of its generality and the fact it is written in JavaScript, the Diagram's AI is quite slow (it searches only 10k positions per sec on a PC, while programs like Fairy-Max or ChessV typically do 500k/sec). So it cannot look very far ahead; adjusting the depth to 4 ply already makes it unbearably slow for large variants. But in a 3-move game thinking 4 ply ahead does not even bring the complete opponent's turn within the horizon. One could save a factor by not searching move sequences that are merely permutations of an already searched line (which for a 3-move game would save a factor 3! = 6 per turn), which would be automatic if the AI used a transposition table. (But the current one doesn't.) And moves cannot always be made in a prescribed order, as they could block each other even when they must be made by different pieces.
Large games often 'factorize' in a number of independent battles, but it is difficult for an AI to exploit that (but easy for a human): even in a single-move game the AI will try to interleave the moves of the varios battles in combinatorially many ways.
I could do a two-move game like Duck Chess (with dedicated code) only because it is not really a 2-move game, but a Refusal variant.
So I guess for now an AI for multi-move games are not really feasible. Even piece drops are too difficult at the current stage, and solving this has a much higher priority.