Unlocking Behavioral Complexity with Biological Optimization Methods
Understanding complex emergent behavior using biologically derived optimization methods
Modern civilizations approach to animal behavior and the natural world comes across as a schizophrenic assortment of opposing sentiments — two parts resource exploitation mixed with a big dose of romantic, enlightenment era notions regarding the nobility of the natural world and how best to preserve it. In this it seems to mirror humankinds’ own confusion regarding our place within the animal kingdom. Had we a better understanding of our unique patterns of behavior and how they fit into the larger picture of evolved decision making systems, I believe we might be better equipped to formulate sensible policies — for ourselves, the ecologies in which we live, and ultimately the machine learning systems to which we are gradually entrusting more and more of our decision making.
Attempts to use data science for analyzing the behavioral complexity of organisms has seemingly arrived at two divergent streams of thought — on the one hand, the neoclassical utility maximizing theory consistent with decreasing marginal returns, and on the other, the increasing returns of path-dependent evolutionary economics. Both have generated supporting evidence from real world historical and behavioral data, however neither sufficiently describes the full picture of behavioral complexity. The following is an attempt to synthesize these two outlooks by means of a third concept — biological optimization methods. This concept provides a bridge between the two apparently contradictory ways of framing behavioral complexity — much the way that unified field theories have attempted to bridge quantum mechanics and general relativity within physics.
While the two aforementioned economic theories have proved effective at gaining insight into certain human and animal behaviors, both have lacked a high fidelity model of behavioral complexity, from the simplest forms of organic life all the way up to to the most complex. The key insight in creating such a model here — the touchstone on which it is based, is that all lifeforms and thus all behavior, in one form or another can be modeled as an emergent optimization process.
Returning to existing methods, it has long been known that the theory of the utility maximizing decision maker posited by neoclassical economics was inconsistent with large swaths of behavioral data. That said, no other model for understanding behavioral complexity has hitherto explained as much of the observational data as this one. Only recently, through the wedding of computer science and behavioral psychology has a more nuanced understanding of behavioral complexity emerged, in a large part based upon what I call biological optimization methods.
One way of looking at the problem is that neoclassical economics started with a top down approach — attempting to explain behavior solely based upon the theory of the rational decision maker. However as we shall see, optimization by so called rational decision making is merely the highest rung in a layer cake of biologically occurring optimization methods, and unless it is considered in concert with the other optimization methods from which it emerged and with which it concurrently functions, any predictions based upon such a model will necessarily be flawed.
Much ink has already been spilled regarding each of the optimization methods discussed in the pages ahead and my purpose is not to revolutionize any one these but rather to connect them in an ordered system, the results of which ought to prove somewhat revolutionary. In short I will attempt to provide a taxonomy of optimization methods at play in the biological world along with an analytical framework for computationally modeling such methods — the Non-Stationary Markov Decision Process(NSMDP). The goal of such an approach is to generate novel data science methodologies for understanding and emulating complex biological behavior. Such techniques should see application in a wide variety of fields including video games, robotics, chat bots, tele-agents and much more.
The four emergent optimization methods I will attempt to classify and connect in an ordered system are — negentropy, evolution by natural selection, reinforcement learning, and causal reasoning. These methods form a kind of layer cake, in which each optimization process supports and interacts with the layers above and below it. Not all biological systems possess this full range of optimization methods, and there maybe others outside my ken. However understanding how these four systems relate and interact with each other should bring us closer to a holistic picture of behavioral complexity and our place within it.
For the computational simulations I will be using MindMaker AI for Unreal Engine 4 — a free open source machine learning toolkit.
Coming Next — Evolution and the Non-Stationary Markov Decision Process
Don’t forget to give us your 👏 !
