Simple Robots Evolve to Become Cooperative

This article was originally posted on RealClearScience.

The evolution of cooperation remains something of a mystery in biology. The cutthroat individualism at the heart of natural selection seems, at first glance, to run counter to the very notion of sociality. However, groups of organisms that display cooperative tendencies tend to be more successful than those that do not, and hence, natural selection also appears to operate at a level higher than merely that of the individual.

Yet, the question lingers of how this is possible. Sociality often involves complex behavior, such as the division of labor observed in ant colonies. How exactly does that evolve? Researchers who have examined the issue tend to assume that the ability to perform each function — for instance, foraging for food or tending to offspring — were present in nonsocial ancestors. According to this paradigm, a solitary insect must be able to do everything, but a social insect will evolve to specialize in just one or a few tasks.

Such hypotheses are difficult to test. But the advent of swarm robotics, which allows researchers to examine the behavior of multiple robots in a controlled environment, has provided insights into evolutionary biology. Now, a team of mostly Belgian scientists has demonstrated, using computer simulated robots, that (1) pre-programmed behaviors are not necessary for division of labor to evolve; and (2) cooperation can evolve entirely on its own.

The authors’ simulated robots mimicked the behavior of leafcutter ants. These ants divide the labor of leafcutting into two different jobs: “droppers” that cut the leaves and allow them to drop onto the ground and “collectors” that collect the leaf fragments and take them back to the nest. There are also “generalists” than can do both. (See figure.)

The scientists assessed the performance of teams of four robots. The robots were pre-programmed to be droppers, collectors, or generalists, and they found that the most efficient teams contained exactly two droppers and two collectors. That is not surprising.

What is surprising is what happened next. The authors used simulated robots that were not pre-programmed with special leaf cutting or collecting functions. Instead, the robots were only programmed to move toward the leaf, move toward the nest, or move randomly. Additionally, they were given the ability to evolve via mutation and crossover followed by selection. (In this case, selection was based on the number of leaf fragments collected.)

At first, the robots hardly collected anything. But selective pressure worked its magic. After 2,000 generations, most of the robots evolved division of labor spontaneously.

Put another way, robots that were programmed only with basic primal instincts and genetic mechanisms, over time and in the presence of selective pressure, evolved to cooperate. That’s pretty impressive.

However, there is one major caveat to the research. The biological relevance of the work largely depends on whether or not their evolutionary algorithm properly resembles reality. Unfortunately, many simplifications had to be made, and the accuracy of various features, such as mutation rates and selection paramaters, are uncertain. Despite that, such systems can provide interesting insights into evolutionary biology, rudimentary though they may be.

Source: Ferrante E, Turgut AE, Duéñez-Guzmán E, Dorigo M, Wenseleers T (2015). “Evolution of Self-Organized Task Specialization in Robot Swarms.” PLoS Comput Biol 11(8): e1004273. doi:10.1371/journal.pcbi.1004273