In most positions, just knowing what moves are legal is enough to give you a good idea of most of the board state.
This is a very good point worth pondering over. I think it applies to many areas. You can see in this video that Carlsen has incredible memory. Even for someone is not Carlsen but has played a long time, simply being able to eliminate invalid board arrangement in less than a second of thinking time is enough to eliminate a lot of extra work that you can spend on less tedious work that require more processing power. I think this applies to musical improvisation too. A lot of good improvisers don’t really need to think about what notes are valid and what aren’t. They’ve practiced so much that it’s second nature.
I feel like neural network are similar in a way that the prior is essentially this second nature memory. Every epoch is the model seeking out new information to add to the existing model and finding out where in the model this new information fits, or maybe even throw it out completely or just in the furthest corners of the model’s memory.
It’s worth noting that in blindfold chess, you by definition don’t see the board state, only the history (if you can remember it), and whether you played a legal move. It has a sort of mirror or dual game in the form of Kriegspiel, you can see your pieces but not the enemy’s moves/history, and you again can try to play moves and will be told if they are legal or not (and you are expected to ‘probe’ with possible moves to gain information from the legality thereof). This demonstrates that human players can play satisfying chess with not much more than legality plus some additional information. (It would be interesting to see if with enough practice, humans could play ‘blindfold Kriegspiel’ reasonably well or if that winds up being too difficult.)
It becomes a probabilistic chess game rather than an expansion of the existing chess game. When you can see the pieces yourself, you don’t have to solve the probabilistic problem. Reducing complexity is a strategy in and of itself. Now you have to weigh whether you want to focus on getting better probabilities, and whether that gets in the way of regular chess strategies.
This is a very good point worth pondering over. I think it applies to many areas. You can see in this video that Carlsen has incredible memory. Even for someone is not Carlsen but has played a long time, simply being able to eliminate invalid board arrangement in less than a second of thinking time is enough to eliminate a lot of extra work that you can spend on less tedious work that require more processing power. I think this applies to musical improvisation too. A lot of good improvisers don’t really need to think about what notes are valid and what aren’t. They’ve practiced so much that it’s second nature.
I feel like neural network are similar in a way that the prior is essentially this second nature memory. Every epoch is the model seeking out new information to add to the existing model and finding out where in the model this new information fits, or maybe even throw it out completely or just in the furthest corners of the model’s memory.
It’s worth noting that in blindfold chess, you by definition don’t see the board state, only the history (if you can remember it), and whether you played a legal move. It has a sort of mirror or dual game in the form of Kriegspiel, you can see your pieces but not the enemy’s moves/history, and you again can try to play moves and will be told if they are legal or not (and you are expected to ‘probe’ with possible moves to gain information from the legality thereof). This demonstrates that human players can play satisfying chess with not much more than legality plus some additional information. (It would be interesting to see if with enough practice, humans could play ‘blindfold Kriegspiel’ reasonably well or if that winds up being too difficult.)
It becomes a probabilistic chess game rather than an expansion of the existing chess game. When you can see the pieces yourself, you don’t have to solve the probabilistic problem. Reducing complexity is a strategy in and of itself. Now you have to weigh whether you want to focus on getting better probabilities, and whether that gets in the way of regular chess strategies.