←back to thread

How to Build a Chess Engine and Fail

(obrhubr.org)
131 points xlinux | 2 comments | 19 Nov 24 06:38 UTC | HN request time: 0.908s | source
Show context
snickerbockers ◴[19 Nov 24 19:05 UTC] No.42186967[source]▶
This is something that's been on my bucket list for a while now, although I'd do it using "normal" programming instead of DL.

i feel like if you can bruteforce every possible board configuration for the next 3 turns and then make the move that leads to more desirable outcomes, that ought to be enough to thrash most amateurs. Probably not good enough to beat somebody who actually understands chess tactics on a deeper level, but I expect most non-masters are just "winging it" and making things up as they go along, so a machine that can think farther ahead than you would win more often than not.

But like I said, this is all just me fantasizing about a project I haven't even started yet so my hypothesis could easily be wrong.

replies(3): >>42187226 #>>42187747 #>>42188789 #
janalsncm ◴[19 Nov 24 22:30 UTC] No.42188789[source]▶
> make the move that leads to more desirable outcomes

This is kind of begging the question, right? How do you know what a “desirable” outcome is, other than mate-in-N positions? (In your case, N<=3.) The point of an evaluation function is to take a position and return a value which describes how desirable it is.

replies(1): >>42189084 #
1. snickerbockers ◴[19 Nov 24 23:14 UTC] No.42189084[source]▶
It's basically just a depth-first tree traversal with some heuristics to weigh different turns against each other (where a turn is defined as a possible board configuration that can be reached from the current configuration).

First you'd have a heuristic that ranks outcomes from turn 3, which would just mean weighing the pieces the computer lost against the pieces the opponent lost. Then you rank each turn-2 node based on how desirable the turn-3 leaf-nodes it leads to are and how likely the opponent is to actually make the moves that lead to it. Then you'd do the same for turn-1 and then pick the move that scored highest.

replies(1): >>42198361 #
2. Quekid5 ◴[20 Nov 24 21:36 UTC] No.42198361[source]▶
Ah, yes... "just". I think you're seriously underestimating the difficulty here.