I guess, someone way find this very old article (written by Alex J. Champandard on February 11, 2008) useful (as I did).

The AI from Half-Life’s SDK in Retrospective

very interesting article indeed, thanks for sharing

Thank you very much 3run!

You are welcome guys! I'm happy, that you've found it useful!

After i read the article i realized something!
IntenseX works the same way!

Intersting article
However it seems to confirm that sophisticated AI tecniques : neural networks, genetic algo , reinforcement learning atc are not applicable in real games
Common sense still dominates

@AlbertoT: The problem with the algorithms you mentioned is that you need lots of training data to get a decent AI. In some cases, like reinforcement learning, it is possible to boost learning by training phases in which the computer plays against himself (followed and preceeded by seesion in which the computer plays against a human being), but this requires more or less that the game is already more or less balanced.

From what I have seen I believe that this is perfectly possible for board games like backgammon and chess; and (team-)sports games, like soccer, icehockey, baseball and tennis.

Though, games tend to be a bit "arcade" like (or in other words: accessible to casual gamers), so virtual characters run faster and jump higher and such. Personal skills play nowadays also a big role in sports games (like in Fifa) and they blast the variables into infinity; I believe that for neural algorithms it is already too hard to capture general strategies and I think it is impossible to capture personalized skills as well (when playing against Real Madrid you try for example to take out Xabi Alonso out of play and treat Christiano Ronald in a different way, compared to your regular soccer team opponent).

Because of these overwhelming set of variables and requirements of the player to a game, tweaking a handcrafted state machine seems to be more convenient to game designers, for which they can try out different things like "lets see how hard the AI is if it can produce Tank Units 50% faster than before" or "lets give the Beast half health points because it is already too hard for the casual gamer". It is for example unclear how Reinforcement Learning encodes the "knowledge" and "experience" in practice and this is a no-go for developing commercial games.

Neural networks are nowhere near to what we need in games, at least how most programmers do it. A lot of tutorials on NN talk about simple multilayer perceptron networks, but it's a lot more than that. The biggest difference between NN and the way biological neurons work is, that it's not a simple output function, based on weighted inputs. Neurons have a fire rate (so to speak), a depletion and regeneration time etc. The architecture of the brain uses a lot of feedback to previous layers, providing iterative results.
NNs are usually the worse choice compared to f.e. behavior trees, but they are able to adapt to the player. For most games adaption isn't really needed because enemies mostly die after a short time. Beat 'em ups on the other hand already use NNs to adapt to combo patterns of the player.

Almost every designer will tell you that it doesn't matter, how smart an enemy is, but how smart he appears to the player. So the best thing you can do is to use the simplest algorithms available that still provide a believable behavior.

edit: have a lookt at this review.

I read the book " Artificial intelligence for games " by Ian Millington
A good book, out of question, but also disappointing up to a certain extent
I would say that 90 % of what explained in the book is common sense
In several occasions, the author makes fun of advanced AI enthusiasts
Those guys have never developed a game, he claims
What about your experience ?
Did anybody design a working NN or genetic or reinforced learning algo, or...?

Definitely true
However if advanced AI algo's should lead to even slight more realistic ( human like ) behaviors then some games would for sure implement them
The point is that it seems that they dont even work
Simple algo's do it better
For example
I designed a fuzzy logic algo
I did not notice any benefit
Ok, I am just an hobbyest programmer I thought
Reading the Millington's book I was surprised to learn that actually nobody can distinguih a fuzzy logic algo from a simple Random.Range() command
Disappointing