Arimaa Forum (http://arimaa.com/arimaa/forum/cgi/YaBB.cgi)
Arimaa >> Bot Development >> "Despondent" bots
(Message started by: aaaa on Mar 22nd, 2007, 12:39pm)

Title: "Despondent" bots
Post by aaaa on Mar 22nd, 2007, 12:39pm
A minor point admittedly, but with many bots I see that if it finds a forced goal for the opponent, it kind of gives this away by playing a nonsensical move often consisting of only a few steps. It might be better to let it behave more gracefully instead and play according to a different optimization (of course as long as this doesn't go at the expense of its general performance). Let's say that after maximizing the minimum number of moves to goal for the opponent, it maximizes its own score as defined by the match rules. I can say, being a poor Arimaa player myself, that seeing the computer aggressively push a rabbit forward is a frightening prospect as it almost invariably means it has a forced goal itself. Since I believe this also would be the behavior if the bot maximizes its score (because the row a rabbit is on is cubed in its calculation), it might entice the lesser player to fall back to defensive play without realizing the forced win for him- or herself, thus possibly giving the bot a slightly better score over the long haul.

Of course, one might rebuke and exclaim the unlikelihood of both reaching a forced-win position against a good bot while not being able to spot the forced goal, in which case I would immediately concede the point, but would still appeal for a graceful losing behavior on account of aesthetics.

I don't know if there are any bots that already function this way.

Title: Re: "Despondent" bots
Post by Fritzlein on Mar 22nd, 2007, 7:04pm
As far as I know, there is only one bot that tries to swindle by playing an optically good move when all is lost: that is Gnobot.  I also know of only one instance in which it worked, i.e. in which Gnobot won from a lost position by making a "reasonable" move, after which the human player overlooked a forced win.  That's not a huge success rate for Gnobot, but it does show that attempting to swindle can pay off in practice.

I suspect that other bot developers haven't followed Gnobot's example because they are more interested in beating other bots than they are interested in beating humans.  Given that opposing bots are not vulnerable to being swindled, it may not seem worth the coding effort.

Title: Re: "Despondent" bots
Post by NIC1138 on May 9th, 2007, 10:35am
Perhaps we should try to make "swindlable" bots??... I've been wonderig for a while what would happen if we make the bots pay more attantion to the more recently moved stones... For this to matter, we would need to measure some kind of dependency between a group of moves, and the distance of the next move. If this exists, this means the game has an "inertia".... some kind of auto-correlation.

Title: Re: "Despondent" bots
Post by ddyer on May 13th, 2007, 4:04pm
It's a common feature of bots that all losses (and all wins)
tend to look alike.  When I really care, I tune the behavior
under these circumstances by adding small tweaks to the
evaluation of wins and losses.  But be very careful.

Title: Re: "Despondent" bots
Post by 722caasi on Sep 11th, 2007, 6:27pm
You could make a loss be less bad if in a great position.

Title: Re: "Despondent" bots
Post by leo on Dec 9th, 2007, 12:31am

on 05/09/07 at 10:35:04, NIC1138 wrote:
Perhaps we should try to make "swindlable" bots??... I've been wonderig for a while what would happen if we make the bots pay more attantion to the more recently moved stones... For this to matter, we would need to measure some kind of dependency between a group of moves, and the distance of the next move. If this exists, this means the game has an "inertia".... some kind of auto-correlation.

Interesting... Surprising and puzzling one's opponent can be an effective tactic at times, although diversion works best in games with no total knowledge of the game state. For a bot to be swindlable, it would need to be able to somewhat model the opponent's plans, ie wondering "what in the world is he after?", otherwise any move is just something that leads to some board state to analyse.

Regarding the "inertia" you suggest, I guess there is some truth in it, but the best players seem to have a good time scattering their move steps all over the board for global strategy and multiple threats, which makes the analysis tricky.



Arimaa Forum » Powered by YaBB 1 Gold - SP 1.3.1!
YaBB © 2000-2003. All Rights Reserved.