This blog is about Artificial Life (Alife) and its related fields: artificial intelligence, evolutionary robotics, dynamical systems, and biologically inspired computing. Alife examines lifelike systems using computer simulations. It assumes that life on Earth is just a subset of a larger concept of what might constitute life in general, or “life as it could be”. Some forms of artificial life may meet the criteria for actual life (can metabolize, grow, respond to stimuli, and reproduce in a virtual environment). The question of whether such systems are actually alive is an ongoing philosophical debate.
Alife’s origins date back to the early 1950s when the field of Cybernetics was born. Cybernetics focused on the interaction between a machine and its environment. The simplest example is a thermostat, a device that uses feedback (thermometer) to alter its behavior (engage heater or not) creating a causal loop. Feedback was a key concept for developing more advanced adaptive machines, but unfortunately Cybernetics died prematurely as it was overshadowed by the advent of the computer.
The field of Artificial Intelligence (AI) was born in the 1956, shortly after the first computers were built. Scientists were very optimistic about their potential and believed it was only a matter of decades before we’d have machines with greater than human intelligence. They were wrong, of course. The goal was to model high-level human reasoning, planning, and problem solving using a top-down approach. That is, break down higher level goals into smaller and smaller steps. This approach works well for very specific tasks, chess for example, but not in situations where the environment changes, or in lower-level cognitive tasks, for example telling the difference between a picture of a cat and a dog, a task easily performed by a 3-year old child. Eventually it became clear, through a series of “AI Winters”, that AI had serious limitations. It is now often referred to as GOFAI (Good Old Fashioned Artificial Intelligence).
AI’s limited top-down approach fueled a revival in the 80s in field of Connectionism, a bottom-up approach that modeled brain processes using artificial neural networks (ANNs). Although ANNs are indeed a vast oversimplification of real brain processes, they are adaptive systems that can successfully identify and classify patterns, e.g. face or handwriting recognition, make predictions given a series of data e.g. in financial markets, and other applications such as controlling autonomous vehicles, and spam filtering. ANNs are now mainly just considered statistical tools, but they are important to Alife research, especially when coupled with artificial Darwinian evolution.
Genetic algorithms (GAs), popularized in the 70s by John Holland, represent the most powerful technique in the Alife toolbox. It is an astonishing fact that evolution is simultaneously a creativity engine capable of generating a potentially unlimited set of novel solutions to a problem, and that it is also relatively easy to program on a computer. Just three ingredients are required: replication, mutation, and selection. At its core, a GA is simply an optimization technique, however it has the advantage of being able to quickly find solutions to problems with a large number of variables, which is generally the case in Alife research. For example, the methodology of Evolutionary Robotics uses GAs to design robot control systems which may involve 100s of design parameters, including the structure of the ANN and possibly the physical dimensions of the robot itself.
The limitations of AI in the 80s caused another revolution led by Rodney Brooks at the MIT AI lab, who realized that intelligence must be built bottom-up by piecing together components in layers. Using this technique, Brooks built insect robots that moved in a lifelike manner, and could handle situations in the environment that it was never programmed to deal with, something that top-down AI could not do. This caused a revolution in the field of AI.
Alife research combines some or all of these ideas: the bottom-up approach, artificial evolution, artificial neural networks, and robotics. In fact, all that is really necessary to demonstrate a bottom-up lifelike process is a simulation of some kind. The first real Alife experiment was John Conway’s “Game of Life” which is a type of cellular automaton, a concept developed in the 40s. Cellular automata are a powerful example of Alife because they show how complexity can emerge from very simple rules.
Alife is fascinating because it may be a method for achieving true artificial intelligence in a bottom-up fashion. It can also inform and validate biological theories. By exploring “life as it could be”, we can learn more about “life as we know it”.
