Perspectives on Intelligent Systems

When I first started my PhD back in 2005 (even my Masters before that), it was not clear to me how the AI I was working on fit into the broader and popular notions of the field. I new it all fit together somehow, although I lacked the perspective or the inclination to piece it together (situation of not seeing the forest for the trees). Two things that occurred to me recently as I outlined my latest understanding of the patchwork of AI related to my work, is the importance of the broader perspective for motivating work, and the apparent lack of papers on there outlining the various perspectives and how they relate. I guess I'm all about abstractions and metaphors so I may have over estimated the topics importance.

I have tackled this problem a few times in the past, the results of which colour my current understandings. Academically, I was introduced to intelligent systems in three ways: (1) first principles such as logic and reasoning, (2) soft computing such as neural and evolutionary computing, and (3) intelligent agents and multiagent systems. A typical formal undergraduate education I suspect. I like the messy stuff over the logic or game theoretic stuff, so that was the perspective I pursued.

In my own time, I ruthlessly pursued my interests which exposed me two different perspectives including (1) game programming such as Bots and non-playable characters for first person shooters, (2) data mining specifically classification and visualisation, and (3) optimisation for programming competitions and related activities. The first and last examples I suspect any game playing programmer type would have investigated.

Early during my project I investigated a perspective that was popular in my research group called Biologically Inspired Computation. Later I investigated a related and more recent perspective called Metaheuristics. As I mentioned, these perspectives, and adaptive systems theory shaped my understanding and therefore the motivations of my own research.

More recently, as I have pulled everything together to paint the picture (write the narrative) of my work for the submission I have realised that all these 'fields' have the same general aim, although importantly emphasised different qualities to promote different (usually distinct) perspectives. For example, Soft Computing was all about less ridged systems compared to GOFAI, Machine Learning focused on pattern recognition and learning paradigms, Biologically Inspired approaches motivated abstractions from the natural world, Computational Intelligence unified disparate strategy-based fields, Agents incorporated rational models, Multiagent Systems unified bottom-up Distributed AI, Metaheuristics integrated Operations Research into the strategy focused approaches, and so on.

The acknowledgement that multiple perspectives exist, and are a good thing, should negate petty field-rivalries. The key for me was to assess the structure of the range of popular AI books, and to integrate the definitions of 'sub fields' drawn from seminal papers. Now the beautiful thing is, I believe the same internalisation could be achieved simply from a comprehensive reading of the overviews provided in Wikipedia. The AI Portal for example is cool.

Anyway, I think the perspective ah-huh moment is critically important to being able to exploit findings and even whole paradigms across sub-fields. This is a huge problem, where, as soon as the nomenclature or terminology is shifted, novelty is (re)discovered. That's cool for the empire builders, although it sucks if you want to want to piece together a bigger picture, or more interestingly have an impact.