Sometimes I wait a too long before doing what I really want to do. I’ve postponed writing this article more than once. And this mirrors the fact that I postponed going to visit John H. Holland until I missed my chance.
Very few academics have influenced my thinking as much as John Holland did. We never met, although I did half my graduate studies in Chicago, only 6 hours away from his home in Michigan. He was actually the person I had most wanted to do a PhD with before I figured that the American system wasn’t for me. When he died, in 2015, I missed my final chance.
I first learned about genetics, when I was about 10, from my dad. He setup and ran the first centre for X-ray Crystallography in Ireland. The idea of coding for anything using a static code of only 4 elements seemed incredible to me. Even more astonishing was the idea that somebody had already done this, with considerable success, on computers.
John Holland was often referred to as the father of genetic algorithms (GA’s). He was one of the first people to take concepts such as genome and selection for fitness and encode them in computer code. I subsequently learned, in personal conversation with Refoyl Finkel when he visited NUI Galway, that the community recognises two independently created strands of thought in genetic algorithms and John was the author of just one of those.
Ever since the early 90’s I have circled back to both this topic and artificial neuronal networks at every opportunity. In the early 2000s I had reached a level of maturity where I could read John’s books. To say that I devoured a few of them would be an understatement. I actually think that I learned my views on how mathematics can, and should, be applied to understanding biological systems from those books.
From my perspective, John specialised in formalisation of biological concepts. He listened to the biologists, wrote down their theories in code, then wrote it in pretty elaborate mathematics. I’ve never had my view independently confirmed – perhaps his sequence was different. He essentially proved his worth as a professor by doing the full mathematical description of his systems, making them tractable to full-scale analysis. That he did this in the most playful and intellectually stimulating way is why I was so drawn to his work.
I never really understood the need to write things out with the full mathematical notation until I came along to academia myself. It is still, from my perspective, not strictly necessary to the type of analysis John typically did. But it is very necessary in order to mark your territory as a professional academic.
Recently, I had a staff member with a background in statistical physics work on developing a new algorithm which strongly resembles John’s schema representations. I was astonished to find out that not only had the employee never heard of John Holland but, when I looked into it, nobody else working in our field seems to have either. The true shame here is that Holland invented a very elegant Message Passing board, which is a rather interesting form of a Turing Machine, with an encoding which I think is very relevant to many modern tasks, but the write-up is largely in books rather than in articles.
John’s work aimed so high, and often accomplished those heights, that he seems to have spent considerably less time than others assuring his legacy. His former postdoc Melanie Mitchell, who I long assumed was the natural inheritor of his work, appears (I am not working in this field) to have shifted her interests to the more popular areas of CNNs and Gaussian Processes. I am not sorry to see a brilliant professor work on these areas. But I am sorry to see one of the people, who I long suspected had the most relevant insights, slowly drifting out of the main thrust of this particular field of research.
I spent some time, two weeks ago, trying to track down a good description of John’s Message Passing / Bulletin boards. They are usually written into the literature as Classifier Systems, but John went far beyond this description in his books. I honestly didn’t find anything that I really want to link to. I had to give up on having a document for my employee. Instead I sketched the idea on a whiteboard… I don’t think that my efforts had the desired effect.
This is the difficulty with truly interdisciplinary work. It is immensely intellectually rewarding. But you have less control over how your work is perceived and adopted. John’s work was strongly adopted by some pretty niche communities. But some other communities, such as my physics graduate employee, could badly do with a primer on these tools. The basic graphical and philosophical approach laid out in John’s work is missing from anything I have found so far online.
John’s work has intersected with my own interests, first with genetic algorithms, then later with complex systems / complexity. His bucket-brigade algorithm is also highly relevant to anybody, including myself, working in associating reward in a learning system.
John spent most of his time since the late 1980’s working on complex systems. I guess that his work was my first introduction to this space. From my perspective, his GA’s are the natural introduction to this field. They have a short, simple list of basic rules but they lead to complex and stochastic outcomes.
More than anything though, I found that I loved his work for its role in my life as a model for how science can be done.