| |
BGonline.org Forums
Is bot analysis of manually altered positions reliable?
Posted By: Timothy Chow In Response To: Is bot analysis of manually altered positions reliable? (Karol Szczerek)
Date: Saturday, 31 October 2015, at 7:52 p.m.
The vast majority of positions that arise are positions that have never arisen in the course of the bot's training. If a bot is going to be any good at all, it must be able to generalize from its experience. Empirically, we find that bots tend to play well, so it follows that it must be doing a decent job of generalizing from experience. This conclusion is backed up by the fact that in most cases, we are able to make sense of the reasons that it makes the plays it does.
It is true the if we set up a position that is very different in nature from anything that it has encountered during its training, or if there's a strategic concept that isn't easy to learn via temporal-difference learning, then we might expect it not to do so well. Crazy superbackgames fall into this category. But your position doesn't strike me as falling into this category.
That said, I don't necessarily trust differences of ±0.020 in this sort of position, but I would not expect the "legal" position's evaluation to be necessarily more reliable than the "barely legal" position.
Anyway, it seems totally reasonable that having a direct 6 to hit is a good thing for a potential blitz, so why would you be skeptical of the verdict here?
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.