Your network blocks the Lichess assets!

lichess.org
Donate

AI Chess Coaches?

@Toadofsky said in #20:

everything's the same old technology with new marketing.

I feel you, man! You get old enough and everything feels recycled. Because it is! :D

But there is a significant difference between '90s tech and today's: the quantity of available data. Crap quality, true, but still a lot of raw material to mine and process. This means new processing methods can break through the plateau you are talking about and this is what we're discussing here. They haven't yet, and with that I agree wholeheartedly.

@Toadofsky said in #20: > everything's the same old technology with new marketing. I feel you, man! You get old enough and everything feels recycled. Because it is! :D But there is a significant difference between '90s tech and today's: the quantity of available data. Crap quality, true, but still a lot of raw material to mine and process. This means new processing methods can break through the plateau you are talking about and this is what we're discussing here. They haven't yet, and with that I agree wholeheartedly.

@TotalNoob69 said in #21:

everything's the same old technology with new marketing.

I feel you, man! You get old enough and everything feels recycled. Because it is! :D

But there is a significant difference between '90s tech and today's: the quantity of available data. Crap quality, true, but still a lot of raw material to mine and process. This means new processing methods can break through the plateau you are talking about and this is what we're discussing here. They haven't yet, and with that I agree wholeheartedly.

Originally I mentioned databases -- but those aren't coaches! I wrote:

Since the 90's there have been no major advances in "AI coaching" unless you count what ChessMaster does. Sure, databases and endgame tablebases are great, but everything's the same old technology with new marketing. But why?

Endgame tablebases with engines have been quite useful, so I guess that counts as a major advancement coaching-related. I guess also Lichess' puzzle generator, like many, are also great advancements, but more importantly it's become much easier to find coaches and to share games and analyses with humans (and to purchase game databases to do your own research using conventional albeit far more powerful tools).

Again I think tailuge's original chess-o-tron (the one which has you identify features whether or not they are relevant to PVs) is brilliant, as much as Maurice Ashley's manually generated exercises are. But aside from engines becoming much, much more efficient there haven't been any creative breakthroughs in the coaching space.

@TotalNoob69 said in #21: > > everything's the same old technology with new marketing. > > I feel you, man! You get old enough and everything feels recycled. Because it is! :D > > But there is a significant difference between '90s tech and today's: the quantity of available data. Crap quality, true, but still a lot of raw material to mine and process. This means new processing methods can break through the plateau you are talking about and this is what we're discussing here. They haven't yet, and with that I agree wholeheartedly. Originally I mentioned databases -- but those aren't coaches! I wrote: > Since the 90's there have been no major advances in "AI coaching" unless you count what ChessMaster does. Sure, databases and endgame tablebases are great, but everything's the same old technology with new marketing. But why? Endgame tablebases with engines have been quite useful, so I guess that counts as a major advancement coaching-related. I guess also Lichess' puzzle generator, like many, are also great advancements, but more importantly it's become much easier to find coaches and to share games and analyses with humans (and to purchase game databases to do your own research using conventional albeit far more powerful tools). Again I think tailuge's original chess-o-tron (the one which has you identify features whether or not they are relevant to PVs) is brilliant, as much as Maurice Ashley's manually generated exercises are. But aside from engines becoming much, much more efficient there haven't been any creative breakthroughs in the coaching space.

@TotalNoob69 said in #11:

This gives me the idea of minimaxing plans! It brings an engine closer to human thought. Basically use the planning DSL as the latent space. Instead of calculating the best move the opponent would make, the question would be what is the most likely plan the opponent is pursuing, and then what is the best answer to that plan. This might even make the need for exit conditions - which was kind of ugly to start with - unnecessary.

@TotalNoob69

I never saw the concept of minimaxing plans before. Your mind marches to a marvel!

I've been thinking about how this may be implemented.

So you want to find the opponents likely plan and assume they will play it and then oppose it. But how to find the plan?

So maybe you need Maia? cos Maia can predict human moves?

But then predicting moves, doesn't tell you about plans. And you still have the AI Slop problem cos AI doesn't get it.

Also a tactic can get in the way of a plan. So you need to integrate a 'plan stopper' which looks for tactical solutions that can halt a plan.

But AI doesn't understand basic concepts tho.

Well thats all I got lol

@TotalNoob69 said in #11: > This gives me the idea of minimaxing plans! It brings an engine closer to human thought. Basically use the planning DSL as the latent space. Instead of calculating the best move the opponent would make, the question would be what is the most likely plan the opponent is pursuing, and then what is the best answer to that plan. This might even make the need for exit conditions - which was kind of ugly to start with - unnecessary. @TotalNoob69 I never saw the concept of minimaxing plans before. Your mind marches to a marvel! I've been thinking about how this may be implemented. So you want to find the opponents likely plan and assume they will play it and then oppose it. But how to find the plan? So maybe you need Maia? cos Maia can predict human moves? But then predicting moves, doesn't tell you about plans. And you still have the AI Slop problem cos AI doesn't get it. Also a tactic can get in the way of a plan. So you need to integrate a 'plan stopper' which looks for tactical solutions that can halt a plan. But AI doesn't understand basic concepts tho. Well thats all I got lol

The plan would be something really generic, like castle long, block the center and attack on the kingside. Bringing pieces to the king side or blocking the king side is a possible anti plan.

After deciding on a plan, a modified SF would add a bonus to any move that follows the plan.

Even if the plan is reevaluated every turn, the existing plan is preferable to any other which isn't much better.

Something like this...

The plan would be something really generic, like castle long, block the center and attack on the kingside. Bringing pieces to the king side or blocking the king side is a possible anti plan. After deciding on a plan, a modified SF would add a bonus to any move that follows the plan. Even if the plan is reevaluated every turn, the existing plan is preferable to any other which isn't much better. Something like this...

@RuyLopez1000 said in #23:

I never saw the concept of minimaxing plans before. Your mind marches to a marvel!

I've been thinking about how this may be implemented.

Yes, this has been on my mind as well, I guess a mixture of using Maia and Stockfish and move down the line to try to heuristically understand what the plan might be, not easy... but worthy of thinking about...

@RuyLopez1000 said in #23: > I never saw the concept of minimaxing plans before. Your mind marches to a marvel! > > I've been thinking about how this may be implemented. Yes, this has been on my mind as well, I guess a mixture of using Maia and Stockfish and move down the line to try to heuristically understand what the plan might be, not easy... but worthy of thinking about...

@TotalNoob69 said in #24:

The plan would be something really generic, like castle long, block the center and attack on the kingside. Bringing pieces to the king side or blocking the king side is a possible anti plan.

After deciding on a plan, a modified SF would add a bonus to any move that follows the plan.

Even if the plan is reevaluated every turn, the existing plan is preferable to any other which isn't much better.

Something like this...

Yes, this has been very interesting to think about, it has kept up up several nights while thinking about this, a Prophylaxis trainer as such, but damn, it is not easy to do...

@TotalNoob69 said in #24: > The plan would be something really generic, like castle long, block the center and attack on the kingside. Bringing pieces to the king side or blocking the king side is a possible anti plan. > > After deciding on a plan, a modified SF would add a bonus to any move that follows the plan. > > Even if the plan is reevaluated every turn, the existing plan is preferable to any other which isn't much better. > > Something like this... Yes, this has been very interesting to think about, it has kept up up several nights while thinking about this, a Prophylaxis trainer as such, but damn, it is not easy to do...