John Holland wrapped up our course, Introduction to Adaptive Systems, with a question answering session for the class. Along with the narrative of his impressive career as the Father of Genetic Algorithms, he related promising areas for research in adaptive systems and then spent a few minutes talking about consciousness.
I’m so glad that I got to take Holland’s class; the complex systems courses I’ve had at Michigan have been more intellectually challenging than anything else I’ve yet experienced. So today, I’m taking a moment to reflect on what I’ve learned from John Holland in a relatively compact summary.
Formal Tools:
- Finding the rules that generate “building blocks” upon which adaptation can improve is very important to understanding complex adaptive systems.
- Default hierarchies leverage homomorphism and parsimony in the representation of transition functions for modeling. Genetic algorithms can build this hierarchy both top-down and bottom-up at the same time.
- A number of different models of interaction can be utilized in examining new research problems; these include billiard-ball mechanics and urns, difference equations, cellular automata, neural networks, classifier systems, and reaction networks.
- Tiered models of interaction are a brave new world that young whippersnappers like us should be pursuing.
Agent-Based Models:
- Performance systems feature conditions or message rules and invoke the power of parallelism.
- Credit assignment for stage-setting actions can be accomplished through a straightforward bucket brigade; credit assignment can also be done by comparing actual fitness values to predicted values and adjusting credit for each rule based on how closely it meets its expected contribution.
- Rule discovery is extremely important to the presence of perpetual novelty that is one of the hallmarks of a complex adaptive system. In Holland’s work, this is accomplished through genetic algorithms, which use building blocks for rules by leveraging the # and * variables with good old binary 0s and 1s. A # means “don’t care,” as in we don’t care whether the value matches 0 or 1. A * means “really don’t care” and means that the value can match 0,1, or #. Rule discovery is more generally achieved through recombination, reproduction, and sometimes replacement.
- Signals and boundaries are key elements in the discussion of interacting agents, such as proteins and food webs. These complex interaction networks have properties of intransitivity, niche roles, phenotypic plasticity, recycling of resources, and diversity of agents.
- Tags are very useful to modeling asymmetrical interactions; by using offense/defense tags, we can model parasitism, trading, and such. Asymmetric mutualism allows modeling of the caterpillar-ant-fly niche roles that are often observed in nature. Selective mating can be accomplished through a mating tag.
- Meta-agents are tiered structures of agents, attached through a tag that determines the strength of adhesion. Induction and competence are features of the agents that comprise meta-agents that relate to how well it can be controlled by other agents or environmental factors. “Seed machines” are one sort of meta-agent we would like to build: imagine if NASA could drop a tiny, simply robot on the moon, and it could grow itself into a mining operation using the resources on hand.
Open areas for research in complex adaptive systems:
- The search for lever points in CAS is ripe for exploration; there are many examples, such as vaccines, but still very little theory. What about lever points in social systems? These include problems of critical mass, communication, and goal alignment in a diverse agent population. Are there modest changes that can shift the direction of the whole system?
- Analogy as a tool for exploring CAS is hardly exhausted; Melanie Mitchell’s Copycat system is an exemplar, but no one is as yet adapting the techniques for problems such as pattern recognition.
- Saccadic recognition of digital images, as opposed to pixel scanning, is a potentially valuable approach to machine perception; John thinks that a bright student could program a model of this in a summer. I wish it were anywhere near my areas of competence so that I could try.
- Tiered models, in which the building blocks at once level can be assembled for building blocks at the next level up, offer a lot of possibilities.
- Language acquisition is an area of particular interest; the days of the universal grammar have passed, and now we seek to understand how we learn language without relying on a built-in set of rules or abilities. The sequencing of utterances is the fuzzy dividing line between animals and people in a way that affects our understanding of consciousness; attaching meaning to information is more than simply reducing uncertainty.