The problem with implementing learning in any game is, ''what do you do if the learning algorithm learns something stupid''?

Most games that use learning turn off the learning code before shipping. They simply use learning to generate interesting behaviours during development. Obvious exceptions to this include Creatures and Black & White.

If you''re going to implement learning, you need to provide some guarantees that you want ''learn yourself into a corner''! How this might be done is an open-ended question.

Timkin

quote: Original post by alexjc
What do you do if your scripter has an infinite loop in his logic?

Sack them.

quote: Original post by alexjc
What do you do when your fuzzy state machine goes into a strange state?

It won''t if you designed it properly. All truth maintenance systems have verifiable output states given input states. Even if you implement a stochastic state machine the distributions over outputs are computable so there is never an excuse for having a FSM, FuSM or SSM with unpredictable outputs.

quote: Original post by alexjc
It''s a shame this unreasonable argument is used to bash learning. Many things can be done to prevent this, just like for the two other problems which developers would now snigger at...

Mmm, I don''t believe that what I wrote was bashing learning. I was simply pointing out a known problem. I am all for embodying game agents with learning and I am well aware of the capabilties of ML systems.

As for the argument being unreasonable, it is not. It is a valid argument given that 99.9% of the gaming industry does not understand machine learning beyond basic classification/pattern recognition using ANNs. Heck, most of the industry still thinks that a GA is a learning algorithm!

It is also a valid argument given the number of clock cycles dedicated to AI. Learning rates for even the simplest systems are of longer time scales than game play. This can be partly accomodated by allowing the system to stabilise to a local maxima of its fitness space before shipping, but it''s very hard to make it stay near a local maxima once you let a savage 16 year old near it! Furthermore, how do you accomodate further learning? During Play? After play? These are all reasonable questions that affect the performance of the game and hence its enjoyability.

On the flip side, I do believe learning can be meaningfully applied to games. Classifier systems are far better suited to most game learning tasks than ANNs, yet almost no one in the industry knows of them (that will change soon). Simulated Annealing and Gibbs Sampling are more appropriate algorithms for many of the tasks that GAs are applied to. Policy iteration and Q-learning can be applied in fantastic ways for learning behaviours in completely unknown environments (where the only way of obtaining information about the environment is by acting and considering the new percepts generated).

If you have ideas for applying learning more effectively to games, then you should do something about it. As they say, the proof is in the pudding. I know that we''d all be very happy to see any positive results.

Cheers,

Timkin