quote:Original post by Russell
The way I look at this situation is that if you are implementing 3x3 tic-tac-toe, and that is the absolute end, then it doesn''t matter. You can hard code everything, and there is absolutely zero point in discussing anything, so let''s just stop if that''s the case.

I believe that was the basis of the original discussion. If you read back through my posts I clearly made the point that we were dealing with a small state space. I''m glad that we agree that such cases are easily hard-coded and that this is a sensible approach.

I''m also happy to extend the discussion to larger state spaces, as I intimated in my last post.

quote:Original post by Russell
I assumed that something more complex than 3x3 tic-tac-toe was the topic, and that a more scalable solution was in order than using a giant if-then-else or any lookup table (fancy or not).

Given your chess programming background, I can understand your bias for tree-based searches and your belief that offline planning is just developing a giant lookup table.

Please refrain from degrading certain techniques by labelling them as ''lookup table (fancy or not)'', simply because in your experience in your specific domain, you haven''t found them appropriate. That doesn''t mean there is no substance to these techniques or that there is nothing to be learnt from them.

quote:
The tree will expand exponentially, and so the nodes you have to search will expand exponentially, and the nodes you have to store will expand exponentially. Unfortunately for your argument, the limits of modern hardware do not allow you to store an exponentially increasing amount of data.

Now here you are either assuming that I am dumb (in that I would try and store an exponential amount of data) or that you are ignorant of current methods in finite horizon planning.

quote:
A depth first search isn''t limited by the hardware.

Not true. Depth first search on an exponentially expanding tree is limited by the hardware to how many branches can be searched in finite time. You''re not going to try and suggest that using depth first search you can search the entire subtree under a given node (to end game) and do this for all subtrees from the current state, are you?

quote:
What you are talking about sounds like something that I have heard people do.

Given your brief description of what your friend does, I''m not talking about storing states encountered during a set of games.

For some reason you are trying to classify techniques as either database techniques or search techniques. What we''re talking about here are planning techniques. ALL planning techniques require search at some stage (or learning if they are learning an agent function via trial & error). We''re not discussing database vs search (at least I''m not, since such a thing doesn''t really exist). We should be talking about offline vs online planning, and/or finite horizon vs infinite horizon planning.

Can we at least come to some agreement about what we are discussing please?

Cheers,

Timkin

PS: My apologies if this post comes across as confrontational. It''s not intended to... I just want to clear up a few things that I believe have been misrepresented.