Erik is right... AI (or rather GOFAI - Good Old Fashioned AI) isn't really Artificial Intelligence. AI algorithms are mostly specialized search algorithms, that more often that not are of a heuristic nature, and may have an ability to "adapt" to varying conditions. In control systems, this "search" can be to look for the proper response to a complex input.

Another way to explain it, AI algorithms involve coding in highly modular and flexible steps so that the sequence and combination of steps can be altered at run-time to adapt to different contingencies based on certain rules.

But they can only produce results based on the search rules explicitly programmed into them, or at most adapt the search rule parameters based on yet more rules... but they cannot generate new Rules based on their experience, which would be true AI. I'd say creativity is the mark of true intelligence (Hey that sounds good!) My theory: Intelligence involves analyzing incoming data to recognize recurrent patterns and to generate rules that can describe and reproduce these patterns. (Whoa, I'm in full-philosophy mode tonight!)

The performance of AI algorithms however is dependent on heuristic and statistical parameters, and often their benefits are only apparent in the long term. So intelligent embedded control systems (which are usually real-time) are a rarity as of now.... Even though I do have a pressure cooker in the kitchen that claims to have a "neuro-fuzzy" embedded controller! :-))

The only real-time AI control algorithms would be Neural Networks. That is because it usually "learns" during the system development phase, rather than during the operation. Once you train a Neural Network (NN) using sample inputs and it is put into operation, the NN can reproduce an output for a given input within real-time. It's fault tolerant nature allows it to give a correct output even to a highly corrupted input (in other words "unpredictable conditions"). Hence there is a lot of application of NN in signal processing.

I am also aware of a research project in intelligent motor control. They are using a Genetic Algorithm (GA) to tune the P,I and D parameters of a conventional microcontroller-based PID algorithm that controls a DC motor. I am not sure, however, whether the GA runs on the micro itself or if the micro communicates with a computer to obtain the results of the GA. However this is not a real-time algorithm, rather the GA will be run for a long time and its results will be hard-coded into the final control system.

The intelligence I can foresee related to embedded systems will come more from complex co-ordination and interaction of simple individual systems, rather than a single monolithic powerful embedded processor. The exceptions will be in real-world applications where control of inconsistent and complex processes is required. Reliability of such systems would be a whole discipline in itself!

Lot of research going on in Intelligent Control and Machine Learning, which are more practical applications of AI. A search on Google on all aforementioned keywords will generate a wealth of informative links. Alternatively you could check out the FAQs on google groups (comp.arch.ai). There is also a lot of good stuff on chipcenter.com.

kundi