Quote:Originally posted by Daerax:
excellent post! one qualm, GA without cross over is not really crippled and is really not so different from Simulated Annealing - the latter of which I have used to tune an SVM with good result

@Daerax:

Off topic, but... why couldn't you just solve a quadratic program? I mean, from my point of view, that's the whole advantage to SVMs; they're built using convex optimization, so you don't need stochastic methods like SA/GA/etc...

Quote:Original post by Steadtler
See kids, this is where GA/DP/ANN leads: semantics arguments over ill-defined methods that gives poor results and takes so much time and hacks and tweaks just to make it work that you could have coded a straightforward solution a hundred times over instead.

IF you don't know what you're doing, sure.

If you know what you're doing, you can get awesome results with a GA on the right problems. Personally, I think GAs are perfectly straightforward.

We don't quibble over the need for or the worth of a decent heuristic for A*, or tell people to stick with DFS because it's more straightforward. So why do we do so over the need for decent operators for GAs? Perhaps it's because we don't have so many concrete applications with concrete heuristics, unlike A* where we have the pathfinding scenario with geographical estimates of distance.

Quote:Original post by willh
I can assure you that GA w/o crossover is very much in the literature; it has been a topic of debate for some time. Maybe you're not reading the right material? Pretty much everything I've written as part of this thread has already been talked about in 'the literature'.

You're insisting on a rigid definition that necessarily includes crossover. I reject this definition while maintaining that the baby is still safely in the proverbial tub. You're also insisting that this view is somehow concrete and the only legitimate one; while that may be true within the GD community, it is certainley not the case!

In your view, what type of crossover is required for something to qualify as a GA? How flexible are you willing to be with your crossover operator while still having, what you would consider, a GA?

In terms of your own GA implementation, what is the goal of crossover?

Originally I believe the argument was crossover is equivalent to mutation. This is correct enough, for a sufficiently general definition of these terms as to offer little information. Any meaningful definition of the terms must partition them. e.g. when considering the mathematical aspects of trying to minimze the probability of destruction of schema, we want to reduce the length of the 'chromosomes' to reduce p_c(S) that crossover will destroy some schema S and reduce the size of the "gene" with regards to mutation affecting S.

Now you are stating that GA does not require crossover. My contention is that while this may make sense GA without crossover loses the main benefit and core idea of what makes a GA: unlike gradient descent (which often takes a long time to converge and gets stuck in valleys) and the like it benefits when you can get good improvements to search by combining good results from different locations.

Without this the GA loses its bite, enough so that it reduces to a poor imitation of an algorithm such as simulated annealing. You may as well use a method such as Best First Search. This is not to say that this method (no crossover) won't work, it will, especially if combining from differing locations is detrimental but it will converge slower than say momentum based methods or SA and may well get stuck at a local extremum.

I think I have exhausted what I can say on the matter. I will leave with this, if you consider the biology metapohor that originated GA then the benefit of crossover vs is mutation is the benefit of sexual vs asexual reproduction. And then this argument reduces to a subject that is older even than computers.

Quote:Original post by Emergent

Quote:Originally posted by Daerax:
excellent post! one qualm, GA without cross over is not really crippled and is really not so different from Simulated Annealing - the latter of which I have used to tune an SVM with good result

@Daerax:

Off topic, but... why couldn't you just solve a quadratic program? I mean, from my point of view, that's the whole advantage to SVMs; they're built using convex optimization, so you don't need stochastic methods like SA/GA/etc...

Excellent question, so maybe I was not clear. We are talking about two different things. Using a stochastic method would be very dumb of me when trying to optimize the margin around the dividing hyperplane but I wasn't talking about this.

See SVM are extremely robust, if you define the problem well and have selected and transformed features appropriately they are very, very hard to beat. They can classify even more complex spaces than a MLP Neural net. But their main weakness is that they don't generalize so well without correct selection of parameters and live and die by correct tuning. Selecting a Kernel function can be done in terms of understanding the problem domain but then after that you have to search for what parameters (e.g. Cost and gamma for a RBF kernel) lead to the best results. This is what I used SA for.

See http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=1333843 for more

I just can't resist the urge.....I have to say something.

There really is no argument when it comes to defining terms like crossover, mutation, selection, etc. because they have all been defined in literature and they all fall under the umbrella of evolutionary computation. There is also an agreed upon definition of an evolutionary algorithm (EA) out there (dating back to 1993).

Just to point out a few things. A GA without crossover is not a GA. It is actually an evolution strategy (ES). And the cool thing about ES is that it varies the amount of mutation based on successes of previously applied mutations. However, later on, crossover was added into ES in one form or another to increase its robustness. Other additions to ES include complex covariance matrices to properly guide search, which at that point, I find it pretty much overkill.

Now as to the merits of crossover, that is arguable, but for all EAs to work properly, it all boils down to exploitation vs exploration. The point of crossover is to promote exploration of the search space by incorporating partial solutions from other candidate solutions. Mutation is simply a form of exploitation of the current solution and local solution space. However, depending on the convergence of the population, the roles of the two can flip around. So, going back to a mutation only EA. To make it work well, you must somehow counterbalance the exploitative nature of mutation with some other mechanism that promotes exploration. In the case of the original ES, it varied the amount of mutation to try and promote exploration, but it seems that comparably, crossover created more exploration, which is why few stuck with the original ES structure.

Back to the original posters question about genetic programming. As cool as that would be to have GP in games, chances are slim that it would become practical. Though there may be some use for it somewhere in the development pipeline. Also, note that GP does not have to evolve actual pieces of code. Most practical applications of GP out there involve using it to do curve fitting by evolving the actual mathematical equation for the curve or data in question. More useful for game AI would probably be Evolutionary Programming, which, if memory serves me correctly, was originally designed to evolve finite state machines, which has a much higher level of relevance when it comes to practical game AI. So, for example, evolving the state machine that defines an AI agent's behavior on the fly through gameplay feedback.

In the past decade or so, people have been trying to improve on EAs by pretty much mixing and matching everything they can come across, so no one really works exclusively with any of the 4 originals anymore (GA, ES, EP, GP). At this point, the lines are so blurred that it is quite pointless in arguing whether something is a GA or not. The fact is that they're all evolutionary algorithms in one shape or form.

Oh yeah, and for people who are obsessed about gradient ascent/descent, there is the Society of Hill-Climbers and the evolutionary society of hill-climbers that is definitely worth checking out.

Quote:Original post by Daerax
[...]search for what parameters (e.g. Cost and gamma for a RBF kernel) lead to the best results. This is what I used SA for.

See http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=1333843 for more

Oh, ok; that makes sense. Have you seen the work that directly optimizes the kernel matrix subject to the constraint that it remain positive definite (e.g. this)? That keeps the problem convex, and is more general than tuning, say, RBF parameters. (Of course, I appreciate that "more general" is a double-edged sword in learning, so I understand why tuning as in your case might be preferred.)

Quote:Original post by noppanit
I'm just curious that has anybody used genetic programming for Game AI before? I mean I'm doing my project and I just need some experience.

I'm not going to get into the genetic programming vs genetic algorithms vs gradient descent argument, but I will say that as I am typing this I am running genetic algorithms to learn strategies for the base set of the card game Dominion: http://en.wikipedia.org/wiki/Dominion_%28card_game%29 . The game has a lot of cards with varying effects and relatively simple effective combos. It would be quite a pain to write intelligent AI for every card, and doing it this way lets me generate multiple AI players that all play differently while still playing fairly well.

In my experience genetic algorithm are not good at finding the "best" answer to a problem, but they are good at finding okay answers. What makes them fun is that they are good at finding unexpected answers.

Quote:Original post by Emergent

Quote:Original post by Daerax
[...]search for what parameters (e.g. Cost and gamma for a RBF kernel) lead to the best results. This is what I used SA for.

See http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=1333843 for more

Oh, ok; that makes sense. Have you seen the work that directly optimizes the kernel matrix subject to the constraint that it remain positive definite (e.g. this)? That keeps the problem convex, and is more general than tuning, say, RBF parameters. (Of course, I appreciate that "more general" is a double-edged sword in learning, so I understand why tuning as in your case might be preferred.)

No I haven't but it looks really interesting. The ability to learn model classes without local minima is great! And the fact they "have also shown how optimizing a linear combination of kernel matrices provides a novel method for fusing heterogeneous data sources" and another paper from MSR I just scanned that tells how to account for missing data in sources is excellent.

But I can't really justify implementing it - it looks like it will take weeks to a month or more of full time work at an unquantifiable gain. I mean the technique itself is more general but there is nothing in the paper that leads to me believe that it learns a more general space. In particular they say its error rate is lower than typical generally but how do I know it is not general only for the specific data they generalized it on? And IMHO the gain improvement is marginal enough to be not worth the effort. Do you have any advice?

See the problem I've faced in implementing a bunch of these advanced methods is that when you are not looking at the iris data set or whether it will rain or amino acids data the methods tend to crumble and barely outperform the simpler methods (out of the box). The methods then aren't so performant (time , space and learning generalization rate) when dealing with fairly large datasets *for a single typical computer* - leading to a lot of time spent "improving by compromising" the method to get better results to your problem. Machine learning is currently part art isn't it?

I've been consider moving a bunch of these matrix based methods to a GPU but my target audience don't have the latest in GPGPUs (i already parallelize so when 8-16 cpus become norm I expect to reap those benefits but with 2 the average the speed gains are not very nice).

Anyways you know alot about this stuff, do you have any specific advice on learning kernels via SDP versus typical methods?

EDIT: To clarify, I am not saying that the more advanced methods aren't worth it. Far from it, they tend to outperform the simpler methods but only after a lot of work done to optimize it to your problem. Where as naive bayes, a decision tree or say kNN just work; Bayesian Nets, SVM or NMF require a significant amount of time 'adjusting' - I find my self reading the proofs of the methods to gain a better understanding for tweaking whereas the other stuff is intuitive - so require alot more work but then blow the simpler ones out of the water. Most of the time. So I always find myself questioning if the improvement will justify the extra work.

[Edited by - Daerax on August 16, 2010 12:17:41 AM]