Dempster was my PhD supervisor and I built on his and Chris Jones's work.
The idea was to use genetic algorithms to find combinations of technical indicators that would turn a profit in high frequency foreign exchange trading. Another PhD student (Yazann Romahi) tried a reinforcement learning approach.
The GA found combinations that worked OK on the training set as long as transaction costs were very low. Outside the training set they sometimes did OK, but, this being genetic algorithms, you could never be sure that it was just luck and over-fitting to the training set. As soon as transaction costs approached realistic levels everything lost money. Another student based his PhD thesis on identifying the 50 successful runs out of 1000 which were significant at the 5% level.
I wouldn't recommend this as a useful approach - the best way to make money in FX is to be a market maker or broker, i.e. make money on the bid/ask spread or to get commission on every trade.
Ultimately I found the mix of finance and artificial intelligence to be a case of the blind leading the blind. The financial people had lots of money and deeply wanted to believe that there were magic patterns in the market that computers could discover. On the other side, the AI people wanted research funding and deeply believed that their magic AI black boxes (neural nets, genetic algorithms, support vector machines, etc.) could discover these patterns. The naivety of and misplaced belief of both sides was quite depressing to observe, although many did (and do) exploit it for their own personal profit - primarily through exorbitant consulting fees.