Author |
Topic: Ability of bots to learn? (Read 740 times) |
|
Belteshazzar
Forum Guru
Arimaa player #5094
Gender: 
Posts: 108
|
 |
Ability of bots to learn?
« on: Apr 22nd, 2010, 11:04am » |
 Quote  Modify
|
I know the subject of creating intelligent bots which learn from their losses has been discussed here, but can any of the current bots do this to any extent? I've been playing Arimaazilla a lot, and find myself going from blowout wins to losses. Is it adjusting to me? Or did I just get lucky with the blowout wins? Or am I trying to be even better than before and end up being worse?
|
|
 IP Logged |
|
|
|
Fritzlein
Forum Guru
Arimaa player #706
Gender:
Posts: 5928
|
|
Re: Ability of bots to learn?
« Reply #1 on: Apr 22nd, 2010, 1:37pm » |
Quote Modify
|
on Apr 22nd, 2010, 11:04am, Belteshazzar wrote:| I know the subject of creating intelligent bots which learn from their losses has been discussed here, but can any of the current bots do this to any extent? I've been playing Arimaazilla a lot, and find myself going from blowout wins to losses. Is it adjusting to me? Or did I just get lucky with the blowout wins? Or am I trying to be even better than before and end up being worse? |
|
Arimaazilla is not learning from playing you. None of the bots learn except for GnoBot, and the server version of GnoBot has that learning turned off. The only way server bots get better is when fast and blitz bots get more CPU cycles due to upgraded hardware, but even changing hardware shouldn't affect fixed-depth searchers like Arimaazilla.
Given that, it is hard to fathom why a bot will play idiotically in one game but come up with crushing moves in the next. I believe the critical factor is whether you manage to steer the game into a type of position that the bot doesn't understand, or instead accidentally stumble into a type of position that the bot does understand.
For example, look at Tuks's three games against marwin in the Arimaa Challenge this year. In one game marwin seemed to be playing quite strongly, and in the other two marwin looked hopeless and lost. The difference was not in marwin, which didn't change at all between games, but in the strategy that Tuks chose. Twice Tuks chose a strategy that marwin couldn't cope with, but once he chose an attack pattern that marwin knew how to exploit and complicate.
This is the fundamental insight behind bot-bashing. If you don't "just play", and instead purposely try to exploit the weakness of the bot you happen to be playing, you can pump your rating far above your level of general skill. Your results against Arimaazilla suggest that you haven't yet identified Arimaazilla's fundamental weakness, or if you have identified it, you aren't playing to exploit it in every single game.
|
| « Last Edit: Apr 22nd, 2010, 1:39pm by Fritzlein » |
IP Logged |
|
|
|
|
Home
Help
Search
Members
Login
Register
Reply
Notify of replies
Send Topic
Print