"What decision do I make now?" : Math will not get you there

[Andrew Prock]

What makes you think that AI isn't math?

- Andrew

[Utah]

It is a silly question of semantics. Please feel free to think of it as math if you prefer [img]/images/graemlins/smile.gif[/img]

[Siegmund]

It's been almost exactly a year now since I saw a post from Utah, when he was first starting this project of his. I have been curious, off and on since then, whether it ever got off the ground or not. Nice to hear it's still going and has produced, well, something, even if we can't see just what.

I am, of course, a bit disturbed by the "math will not get you there" phrasing. I would word it more like this: given your knowledge about the cards and your opponents, there is a straightforward mathematical solution - calculate EVs, pick the best one. The problem is that the kind of knowledge we gain by watching an opponent (especially one who doesn't go to showdown) is difficult to represent in the kind of equation you saw in high school.

I am burning with curiosity about how you train a neural net without feeding it historical information, though. Updating the weights on nodes is not fundamentally different from updating any other statistic based on observations as they come in...

[LuckyTxGuy]

First off let me say that this thread leaves me pretty confused! [img]/images/graemlins/crazy.gif[/img] Utah, I understand that you are developing a new product (software?) that cannot be exposed yet. However, can you give us any idea of if and when this new product might be released for sale or viewing to the public?

[Utah]

[ QUOTE ]
First off let me say that this thread leaves me pretty confused! Utah, I understand that you are developing a new product (software?) that cannot be exposed yet. However, can you give us any idea of if and when this new product might be released for sale or viewing to the public?

[/ QUOTE ]Do know yet. The product is finished for the most part and it can be fully used in gameplay now. I am meeting with the development and investors in about 2 weeks to determine next steps.

[Utah]

[ QUOTE ]
I am burning with curiosity about how you train a neural net without feeding it historical information, though.

[/ QUOTE ]You cant. One of the biggest problems with nets is that you often need lots of data. What makes you think that we dont train it with historical data? What if we have all the data or can get all the data from every hand played online at any site. What if we could train the nets in realtime and what if we could run those nets in realtime firing at the rate of 65,000 nets in 1/500 of a second. That would make the software slightly more powerful than pokertracker [img]/images/graemlins/smile.gif[/img]

[toms2866]

The AI approach is extremely interesting. How would you compare your approach to Poki? Specialized math-based bots like Vexbot have proven to be extremely good hu - are you focusing on fr, sh, sng, multi-table tourneys? Limit, pl or nl?

[Utah]

I remember being very scared by Poki when we were first raising funds as there was no reason to build something if it has already been built. Since we were backed by very serious money we even thought of trying to buy them, partner with them, or license from them.

However, after further analysis it became apparent that Poki suffers from 2 huge design flaws/limitations (cant say what they are). I was sure we could build something better so we simply decided to ignore them and move on. I am sure Poki is very good. However, I am sure that if you put Poki vs us against unsuspecting players we would seriously outperform them.

I know nothing of vexbot. Also, we are not a bot. We went down that path at one point and we have the schematics to build it. However, we wanted to create something commerically available and we didnt want to cross ethical lines.

The problem is that what we have now is way to powerful for public sale. So, we are a bit stuck at the moment. We might dumb it down significantly as to make it acceptable to the poker community - but of course that isnt very fun to do [img]/images/graemlins/smile.gif[/img]

[LuckyTxGuy]

Thanks for the info. I guess there is one thing that sort of leaves me puzzled...ok it all leaves me puzzled.lol But if this application is as phenomenal as it sounds, what happens if a full table of players are using it at the same time? There is only one winner in every hand. If everyone at a table was using it, would it then negate it's usefullness? (If that makes sense)

[Siegmund]

[ QUOTE ]
[ QUOTE ]
I am burning with curiosity about how you train a neural net without feeding it historical information, though.

[/ QUOTE ]You cant. One of the biggest problems with nets is that you often need lots of data. What makes you think that we dont train it with historical data?

[/ QUOTE ]

I was quite sure you did. I was being a smartass - in reference to the comment that "you don't care about historical win/loss" when, in a sense, that is the ONLY thing a neural net cares about, since the network learns by strengthening the connections that lead to winning predictions and weakening the ones that lead to losing ones. (If you watched many tables, extending the idea of "winning" to be correctly guessing what should have been done at tables you only observed.)

Anyway, will be interested to see what comes of it, as well as interested in the more general question of whether a neural network method will outperform other methods attacking poker problems.