Back in 2006, three years in to my degree in “computer systems and software engineering” at the University of York, I embarked on my Masters project. The ambition was quite clear – to use a complementary suite of AI algorithms to simulate speciation. The choice of algorithm for behaviour selection, Rodney Brooks’ subsumption architecture, was dictated by my supervisor, but despite the existence of prior work on the use of the subsumption architecture, the existing implementation was missing key prerequisites for achieving the ambition of the project. Specifically the ability to vary the geographic (and other environmental) conditions such that speciation might occur.
Whilst I learned many things from the arduous process of trying to run successful experiments on this journey, none were so precious as the respect I developed for the complexity of systems. The combination of two core AI algorithms, a broad range of environmental variables and the entropy of unpredictable individual behaviour across a population of hundreds over thousands of generations, led to a huge degree of complex, emergent outcomes. Controlling these outcomes became a huge challenge for me, preventing me from obtaining the evidence I needed to support some conclusions I knew it was possible to draw. The experience was humbling, educational and frustrating in equal parts. The write up I submitted can be viewed at this link. I will eventually edit and post a subsequent revision which includes conclusions drawn after further experimentation past the submission deadline!