Today at ALife was OEE2 – the second Open Ended Evolution workshop. There were 12 presentations followed by a discussion.
I got to go first (which is always the best, as I can then relax and concentrate on the rest of the talks). I was presenting our recent work on defining open-endedness by giving a definition of novelty in terms of the model and meta-model of the system being observed. It turns out that there may be a connection between this definition and some definitions of creativity. I’ll have to chase up some more references.
There were some interesting themes running through the day. I’ll pick out just a few that I found resonated particularly well with me. One theme was on infinite (or maybe better unbounded) scalability: make sure there are no limits designed into the system, because they will, sooner or later, stop open endedness. There were also several talks on using discrete dynamical systems as the basis for either theoretical or experimental investigations. One of these, given by Alyssa Adams, used coupled elementary cellular automata to investigate how using an “environmental” CA (E) to change the “organism” CA (O) rules could result in O displaying certain kinds of novel behaviour. Particularly interesting was that just allowing E to override O’s rules, or overriding them randomly, wasn’t that interesting, but allowing a combination of E and O’s current state to override the rules gave the interesting novel behaviours. A very elegant idea. I think it would be interesting to allow the higher level rule that changes the CA rule also to be changeable (maybe it already is; I need to read the paper). And connecting to the unbounded scalability theme: maybe there should be a rule governing the size of the CA to change, so that the size of its state space can grow unboundedly (although would then make the character of the current analysis somewhat different).
Nathaniel Virgo gave a talk about a modification to Fontana and Buss’ lambda-calculus artificial chemistry. The original in irreversible, and collapses to a boring system unless effort is made to exclude certain kinds of reactions. Nathaniel added an extra kind of reaction, making the overall system reversible (in that it doesn’t lose information about the kinds of particles originally present). This seems to have solved the boringness problem in a straightforward manner. Not only does the new system build interesting reaction networks, it builds interestingly different ones each run, rather than having some average behaviour. The intuition is that the reversibility allows the system to “backtrack” if it gets stuck somewhere, and explore a different path. Presumably the fact that the state space is so huge means that the new path can be significantly different in character form the old path. To start with, any approach that does not destroy information might be fine, but eventually, in order to build in an analogue of thermodynamics and Gibbs Free Energy, reversibility will be required.
The presentations were rounded off with a thought-provoking one by Ken Stanley. He was arguing that OEE systems need to be “interesting”, and that “interestingness” is subjective. As an example, he asked us to consider a though experiment. Suppose there was some agreed-upon complexity metric C that we could use to determine how complex or whatever a particular string is. Then do a random, or exhaustive, search over strings of length M to find the best on. Then repeat for strings of length M+1, M+2, …. The result will be a sequence of the best strings in ascending order of length: a form of open-endedness in some definitions. But would that sequence be interesting? He claimed not, and that there also needed to be some narrative about the process of discovering those strings to make the result interesting (to us). Some in the audience disagreed: they would be interested in those strings! Let’s think of something like a complex work of art. We might find it intrinsically interesting, knowing nothing about how it was produce. We might find it more interesting if we also knew the history: but that’s a “bigger” system, art work plus historical context, so it has the opportunity to be more interesting. Additionally, there’s the interestingness of the metric itself: how was that decided upon? Is it even computable? Then there’s the sheer scale of the problem: exhaustive search quickly runs afoul of combinatorial explosion: the process would never actually work. The narrative of how the strings were found in an evolutionary or other manner is interesting partly because it tells us how this combinatorial explosion of exhaustive search was avoided in this case.
The day finished off with some summary and discussion,and we all went away with our heads buzzing with new ideas, and old ideas looked at in a new way.