Author |
Topic: Research hints (Read 784 times) |
|
NIC1138
Forum Guru
Arimaa player #65536
Gender: 
Posts: 149
|
 |
Research hints
« on: Aug 2nd, 2007, 6:42pm » |
 Quote  Modify
|
Hi there, folks!...
I've been away for a while, finishing my master dissertation... Now it's done, and I must come up for a proposal for a doctorate research!...
I know there are some researchers around here, or at least people who like to read about the subject... So I decided to ask you guys you current view on research on artificial intelligence and arimaa!... and of course, "real intelligence"!...
To my doctorate subject, I've been willing to try out something related to Arimaa, and I wanted to know if anybody out there have suggestions on nice things to try out.
What I really want to study is computation in general, and artificial intelligence... More then that, intelligence itself, the essence!... 
I've taken a look upon some stuff like the cognitive architectures, and programs like EMI and Copycat, and I have read a bit of Minsky, Maturana, Robert Rosen, Hofstadter, Pylyshyn and others. I believe all authors of "these" topics agree one time or another on the importance of hierarchy in understanding the mind and nature (understanding "unedrstanding"!...  )
So, I wanted to do something like comparing game-playing programs, or different computational-intelligence systems, and see how they re oganized. How their internl hierrchy of little computational components work and deal ith each other. And of course, how the system deals with its environment, and how is the hierarchy of the environment itself.
All this is more like a motivation... What I will probably do is just looking into similar programs that solve similar problems.
The teacher I want to study with studies Markov Decision Processes, and Reinforcement Learning, and that will definitely be the place I will start my random walk on the frontiers of science trying to stumble upon something interestng.
Anyway, I realy wanted to try to use arimaa as a kind of benchmark, a problem to apply different learning techniques and see what happens.
Of course the whole of arimaa would be probably too much, but I believe we can imagine a lot of different small problems that feels like arimaa... I'm thinking something like mixing up arimaa with sokoban and catch-the-flag. all of that looking at it as a two-dimensional cellular automata!! =)
So, how crackpot am I? What is the chance I can bea accepted at a university showing this message to the comission that will select the candidates fo the program??... 
|
|
 IP Logged |
|
|
|
aaaa
Forum Guru
Arimaa player #958
Posts: 768
|
|
Re: Research hints
« Reply #1 on: Aug 2nd, 2007, 8:59pm » |
Quote Modify
|
Well, recently some encouraging results have come from computer go programs that employ UCT techniques. Since then I've been wondering whether Arimaa would also lend itself well for such an approach.
|
|
IP Logged |
|
|
|
Fritzlein
Forum Guru
Arimaa player #706
Gender:
Posts: 5928
|
|
Re: Research hints
« Reply #2 on: Sep 4th, 2007, 4:47pm » |
Quote Modify
|
I don't think I have anything to suggest on the level of abstraction you are proposing. My interest is in a somewhat different vein than yours: I wonder what it takes for a computer to play Arimaa well, while you wonder what is the nature of intelligence per se.
However, I recall you mentioning that you might have access to a supercomputer. There are some narrower questions that you could study in conjunction with Arimaa, along the lines of how much it helps to have extra computing power.
I did a little experiment where JDB provided me with a bot, and I had the bot play itself where one side got to think twice as long as the other side. It seemed that the side with doubled thinking time was about 60 points stronger. However, my computer wasn't fast enough to run a large enough number of trials for statistical confidence. With a supercomputer, a more accurate experiment could be run in a minute.
More recently I was looking at various versions of the same bot playing each other. That made it look like each extra step of search depth is worth about 100 rating points. If this formula holds true, BombP4 should be about 800 points better than BombP2, i.e. it should play at World Champion level. But sadly we have no computer fast enough to run BombP4, except postally, and even then Bomb sometimes takes more than a week to move. A question very relevant to the Arimaa Challenge is whether Bomb on a supercomputer could beat the best human in live games.
A third, slightly broader, issue is the difference between execution and training of a learning system. It might be totally beyond the power of a regular PC to play the millions of games necessary to train its own intelligence. However, if a supercomputer played all the training games (an activity you can easily parallelize), an ordinary PC could execute the trained system. This would be a legitimate way to win the Arimaa Challenge! The super-computer would not be playing the actual games per se, it would just be doing the grunt-work to develop a highly-tuned engine, which could then run on a regular PC and beat all the humans.
Just a few ideas on how to use academic computing power...
|
|
IP Logged |
|
|
|
|
Home
Help
Search
Members
Login
Register
Reply
Notify of replies
Send Topic
Print