The Evolution of Robots: Natural Selection in Action
Phil Torres
2012-07-24 00:00:00



What is Evolution?

(people already familiar with Darwinian evolution can skip ahead to the second section on evolutionary robotics)

Before discussing natural selection in detail, we should make a few basic distinctions. To begin, there’s an important difference between the factand the theory of evolution. A fact is a state of affairs in the world, whereas a theory is a set of propositions or a description of a mechanism that explains or predicts a state of affairs.

Explaining and predicting phenomena is the whole point of a theory. Why care about explanation and prediction? Because, on the practical side, being able to explain/predict phenomena enables us to control the world in ways that are desirable to us; on the theoretical side, scientific explanations provide deep understanding of phenomena, and understanding is something worth having in and of itself. (Thus, the scientific enterprise has a “pure” and “applied” side: the first strives for knowledge about what reality is like and how it works, and the second aims to manipulate reality for the purpose of making it better suit our needs. I’ll say more about issues like these in Chapter 17.)

Two questions arise from the fact-theory distinction. First, there’s the factual question “Did X actually occur?,” where X is a purported event in the world; and second, there’s the explanatory question “Why or how did X occur?” The second question is obviously not applicable if the first is negative – one doesn’t need to explain why the sun collided with the moon if the sun didn’t actually collide with the moon! […]

Applying this to evolution, then, we have the initial question “Did evolution actually occur?” and the follow-up question “If so, why or how did evolution occur, i.e., what is the causal mechanism behind evolutionary change?” Historically speaking, Darwin’s initial contribution was to convince the scientific community that evolution is a fact about the biological history of life on Earth – that all living organisms descended from a common ancestor. And he did this by adducing a mountain of evidence in his 1859 publication. [...]

When biologists talk about evolution, they’re talking about changes in the frequency of a trait (or characteristic, or feature) within a population of organisms. For example, if a population goes from 100% of its members having striped patterns on their bodies to 50% of them having dotted patterns, evolution has occurred. It follows that evolution is a population-level phenomenon – i.e., it’s not something that happens to individuals themselves but to groups of individuals. Individuals develop, populations evolve. (A species, by the way, is just the largest possible population of a potentially interbreeding organism.) This being the case, evolution is also a transgenerational phenomenon – i.e., it operates over many generations of individuals, and because of this it occurs very slowly. This relates to a feature of evolution that scientists call gradualism. It explains why one can’t observe the evolution of one’s own species happening in a single lifetime.

Evolution can have many different causes. The mechanism of genetic drift, for example, results in random changes in the frequency of a trait that are neutral with respect to an organism’s reproductive success (i.e., fitness). Natural selection, on the other hand, specifically results in adaptive evolution. Here’s one way to understand the situation: think of organisms as having features and environments as composed of factors. A feature could be any aspect of an organism’s anatomy, physiology or behavior (its phenotype) and a factor could be either biotic (living) or abiotic (non-living). For a species to survive, its features must adequately complement the relevant factors of the environment. When the environment changes – and environments are constantly in flux – a mismatch between the organism’s features and the environment’s factors can result. This is when natural selection jumps into action and attempts to restore the lost feature-factor complementarity. (See Figure 1.) The resulting changes to an organism’s phenotype are precisely what we recognize as the design properties of biology. An adapted organism is one that matches its environment like a key fits into a lock, or like a car fits to the road.

The mechanism of natural selection itself consists of three simple conditions. These conditions are individually necessary and collectively sufficient for natural selection to occur. This means that without any one of these conditions, natural selection cannot happen; and as soon as all three are satisfied, natural selection is guaranteed to occur. Basically, natural selection is like a lever: if one end of the lever is pulled, then the other end will lift the rock that it’s under.



Figure 1 (right) : “O” stands for “organism” and “E” for “environment.”

Condition 1: variation in the population. That is to say, there must be differences in the phenotypes, or observable features, of individuals. Condition 2: differential reproduction. This means that there must be differences in the rate of reproduction (or fitness) of individuals in the population. For example, if an organism that lives in trees has poor balance (a phenotypic feature), then that creature is not likely to survive long enough to leave many offspring. In contrast, if an organism of the same sort has exceptional balance, it will probably reproduce more. But none of this would matter without Condition 3: heritability. That is to say, there must be some way for the phenotypic features of parents to be passed down to their children. If the exceptional balance of an arboreal critter can’t be inherited by its offspring, then that adaptation won’t live on and the population won’t evolve to have better balance. This is the heart and soul of Darwin’s mechanism. As I like to say, the theory of natural selection isn’t rocket science.

[...]

The Evolution of Robots: Natural Selection in Action

But how do we know that natural selection actually works? Have people ever observed natural selection in action? The answer is yes! For example, scientists working on the Galapagos Islands (where Darwin devised his theory) have observed the beaks of finches evolving through natural selection, as a result of environmental pressures brought about by droughts and other such phenomena. But one doesn’t need to travel around the world to see the efficacy of Darwin’s mechanism. One can actually use natural selection (in the form of “genetic algorithms”) to solve immensely difficult optimization and search problems. And one can implement natural selection in populations of robots to create artificial beings that exhibit a kind of proto-intelligence. This is, in fact, precisely what researchers in the exciting new field of evolutionary robotics are doing. The present chapter will discuss some of their findings.

To begin, computer scientists tried for decades to directly program computers to have minds like our own. Although expectations were once high, the direct programming approach to AI (artificial intelligence) has repeatedly failed. This led some theorists to think: “Although we haven’t succeeded in designing a being with an intelligent mind, natural selection definitely has – after all, here we are. Thus, why not get natural selection to program an AI for us? Maybe the problem of creating intelligence is so difficult that – ironically – only a really ‘dumb’ mechanism like natural selection can solve it.”

Evolutionary robotics has yielded some amazing results. Consider an experiment – not the most impressive, by the way, but one that doesn’t require too much background knowledge to understand – that was conducted by Dario Floreano at the Ecole Polytechnique Federale de Lausanne in Switzerland (one of the best technical institutions in the world). This experiment used a type of robot called Khepera. It has a circular body supported by two wheels and eight light sensors to help it navigate its environment: six are located at one end and two at the other. Floreano created a square pen for the Khepera robots to roam around. In one corner of the pen was a black quarter circle that represented a charging station. Each robot was given a battery life of only twenty seconds – that is, unless it rolled over the charging station and got another twenty seconds. The robot could thus potentially live forever, just as long as it kept charging up before its battery died out. (See Figure 2.)



Figure 2 (left)

Here’s how the experiment worked: each robot was controlled by a simple artificial neural network. This “brain” consisted of a set of artificial neurons (or nodes) that were connected together so as to mimic the structure and function of biological neurons. Thus, the light sensors would send signals (containing information about the robot’s environment) into the network. These signals would make their way through the network, but how exactly they’d do this would depend on the connection strengths between particular neurons. The output signal would then travel to the robot’s wheels, where it would control the robot’s movement. In this way, the robot would convert sensory information into behavioral actions. (E.g., “There’s a wall in front of me, so turn left.”)

This being said, the Khepera robots also carried around with them a set of genes. These genes were, basically, the instructions for how to build a particular artificial neural network – that is, a network whose nodes are fixed but whose connection strengths are variable. Consequently, the genes of a robot determine how it will respond to specific stimuli from the environment. While one robot might turn to the left when it approaches a wall, another might turn to the right. These are phenotypic differences controlled by the robot’s brain and encoded by its genes.

With all this in place, Floreano created an initial population of robots and put them in the pen. The best adapted robots – by stipulation – were those that moved as much as possible and stayed away from obstacles; this was Floreano’s “fitness function.” At first, the robots moved about randomly, and many failed to recharge their batteries. But a few rolled over the charging station by chance. These were deemed more adapted than the others because they had twenty extra seconds of life. Floreano then created a subsequent generation from these robots plus some random genetic mutations (to increase variation within the population). The result was a bunch of new variants whose features were derived in part from what had previously worked best. Floreano kept repeating this until, after 240 generations, the Khepera robots were able to move about their environment without bumping into anything and – I kid you not – dart towards the charging station a mere two seconds before their batteries died out. In other words, through a “dumb,” mechanical process, the robots evolved a kind of quasi-intelligence. What’s even more amazing is that the robots’ artificial neural networks evolved such that some neurons would only fire when the robot occupied a specific place in the pen; other neurons would only fire when the robot was oriented in a particular way. This is extraordinary because one finds neurons exactly like these in the brains of biological organisms. The rat brain, for example, also has “place cells” and “orientation cells.”

Thus, the Khepera robots acquired not only the ability to avoid obstacles and recharge their batteries, but an artificial brain with some of the same kinds of cells found in biological brains. In this way, Floreano’s experiment demonstrates just how powerful the mechanism of natural selection can be – it shows that, given enough time, intelligent beings can arise via an entirely unintelligent process. The Khepera robots evolved over a period of two weeks; life on Earth has been evolving for 3.5 billion years.