Author |
Topic: Evaluation MindMap (Read 3871 times) |
|
ntroncos
Forum Junior Member
Weiser's trainer
Gender: 
Posts: 9
|
 |
Re: Evaluation MindMap
« Reply #15 on: Apr 13th, 2006, 10:44am » |
 Quote  Modify
|
Some of the aspect discussed in this thread could be achieved in a way by a double evaluation.
First do a shallow search to look for the possible moves, then line them up in a priority cue.
Then a second in more depth evaluation for the higher priority moves.
|
|
 IP Logged |
|
|
|
99of9
Forum Guru
Gnobby's creator (player #314)
 
Gender:
Posts: 1413
|
|
Re: Evaluation MindMap
« Reply #16 on: Jul 20th, 2006, 10:13pm » |
Quote Modify
|
Just an update on Gnobot's evaluation overhaul:
I've written about 15 functions out of about 50 that are now on my map.
(*) means some subfunctions are missing, but the formula for combining the subfunctions is written.
* game stage
* board balance
* key positions held
* pieces organized
* board structure*
* DAPE - depreciated arimaa piece evaluator
* DAPE derivatives
* Path weights to centre
* E access to centre
* M_E separation
* initiative*
* general advance
* long term win probability*
* short term win probability*
* overall evaluation*
So far these fragments take 1100 lines of code, and I probably subconsciously chose the easiest ones first. Another problem is that as I think more, I add more components to my To Do list. So there's a long way to go.
But at least I think I'm doing each component "properly", and I hope I can get the whole thing ready before this year's CC.
The downside of focussing so much energy on Eval is that I'm unlikely to change the search routines at all - so Gnobby will be a 10 -11ply bot once again.
|
|
IP Logged |
|
|
|
Fritzlein
Forum Guru
Arimaa player #706
Gender:
Posts: 5928
|
|
Re: Evaluation MindMap
« Reply #17 on: Jul 21st, 2006, 10:41am » |
Quote Modify
|
I'm thrilled to see Gnobot get an evaluation overhaul. It was already more strategically aware than the other bots, so this could potentially vault Gnobot into contention for the Computer Championship.
My understanding is that for computer chess, most of the progress came by improving search rather than by improving eval, but I don't see why that would necessarily apply to Arimaa. My hunch is that bots can do much better with static eval than they do at present.
|
|
IP Logged |
|
|
|
|
Home
Help
Search
Members
Login
Register
Reply
Notify of replies
Send Topic
Print