2
$\begingroup$

I had an idea to use the this algorithm to simulate real life game scenarios where a player could train and retrain their mistake and be ready for when that situation happens again in real life. This simulation would calculate every possible scenario and give you the best one. This would work almost the same way as it is used on Stockfish for chess. It calculates every possible scenario and always makes the right move.

Now my question is, would that be possible? Is there another algorithm/AI/Software that could make this simulation work?

$\endgroup$

2 Answers 2

3
$\begingroup$

It also depends on the type of game.

The problem with Go is that a board that looks really good can turn into a disaster on the next move.

Games that have easy evaluations and experience only small incremental changes with each move are much better suited for alpha-beta pruning.

For instance, Checkers could be trivially evaluated as the difference in number of pieces each player has. Most moves will change the value by 0 or 1 points, with double jumps being much less common, and higher numbers being rare. Not going just one level deeper is usually not likely to be a significant problem.

$\endgroup$
2
  • $\begingroup$ Checkers could be trivially evaluated as the difference in number of pieces each player has that's a pretty greedy policy... you probably need something more than that $\endgroup$
    – Alberto
    Nov 21, 2023 at 22:27
  • $\begingroup$ @Alberto, I said it "could" be, not because it would be the best way, but because calculating the value would be very efficient. For Go, there is no evaluation even remotely that efficient, which illustrates the fundamental difference that makes alpha-beta appropriate for one but not for the other. $\endgroup$ Nov 22, 2023 at 0:26
0
$\begingroup$

You can do that, however you're probably going to hit the wall that alpha-beta pruning got with Go and large scale games... it does not scale too well

Depending on what's your definition of when that situation happens again in real life you might need very large and deep trees to get a good estimate, let alone that to work well you have to have a good bounding algorithm, which is not trivial to obtain

As Neil, you probably also need to develop a very sophisticated and complex model to simulate your scenario

$\endgroup$
1
  • 1
    $\begingroup$ For "real life", the OP would also need a rather sophisticated predictive model for exploring "every possible scenario". $\endgroup$ Nov 21, 2023 at 19:51

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .