harrytrader,
I totally agree with you that AI was hyped as a term, and I remember how it was almost taboo in the late 1980's. Still there were persons continuing doing work in the field, and that's why we have more diversified options today, like Kohonen maps etc.
I also use my own neurons to do my trading - and will continue to let them have the last say. Still, I find that I do trading based on not an infinite state of mind, nor infinite variations of signals. Sometimes, however I lose my concentration, and do not process those same signals as well as I do other days. Therefore I am really looking for some confirmation and continuation in identifying some of the indications I do on a daily basis. Thus a more sophisticated means of compiling indicators/signals would help me, if I can have some of them modeled precisely. Then there are other contra-indicators and signals that I would use to make trading decisions.
When I use some pattern for basis on my trading decisions, I sometimes have a very short timespan to close the trade, and these times I would rather have an automated trigger help me, because I know I sometimes am not focused enough to consistently adhere to my current system/rules.
I imagine some mixed system for this, using rule based logic, and then more probabilistic learning methods to try and model some of the indicators I (think I) "know" already. Rule based logic can be programmed in procedural languages, but much more elegantly in declarative languages. It's just like Lisb and other functional programming languages is way more efficient at e.g. list processing like time series etc. I remember a language called Clean, developed by Univeristy of Nimegen (kind of Miranda-like, Haskell) with lazy typing and quite efficient performance.
Those kinds of ease-of-expression tools/languages does help me model things more efficiently, and therefore I look at things like Prolog etc. for easier prototyping - just like I used Beanshell to very quickly deisgn my current trading/analysis platform.
Now, it seems that HMMs are quite good at recognizing sequences of states, but won't necessarily generalize well. This might be a good thing if the markets are constantly changing, in my opinion. Likewise, I have seen promising features of Bayesian Networks - as well as my new found interest in these kinds of networks - which I'd like to explore some. They seem to model well on high frequency data, unlike NNs which would prefer very normalized data, with extremes filtered out etc. NNs are better for generalizing, although I am reading how BNs and HMMs also can be made into generalizing by using median terms.
If one used one's own trading decisions as one of the inputs, how would this possibly map out in some of the models .. could the model quite accurately map your decisions over some timeframe. With "mood" data also entered, would it be easier (nervous day, anticipation, bearish, bullish etc. based on news and other data) ?
Could a model advise you when not (!) to do a trade based on previous failures ? It's interesting, and probably some highly individual results could be obtained.
With regards to Kohonen maps (SOM), I'll look into that too eventually with the JOONE library -
http://www.jooneworld.com , and see if a model for me is feasible with this library.
Besides, all the software and libraries I am using are free.
. 
