For those who enjoy the integration of biology and technology, a company based in Switzerland is demonstrating and studying the effects of robotic personality genomes. Flipping the normal order, by improving technology with biology, engineers from the University of Lausanne, Switzerland are using a program-perfecting method as old as life itself.
Darwinism, the messy, trial-and-error, results-focused system of survival of the fittest can be used to program robots; no intelligence required. The researchers at the University of Lausanne, engineers Sara Mitri, and Dario Floreano, and evolutionary biologist Laurent Keller, have built small robot platforms, with modular sensors, and a special processor. Instead of a traditional number-crunching logic machine, the ‘brain’ of these robots consists of ‘neurons’ attached to the sensors and their wheel-tracks.
The robots are then assigned ‘tasks,’ where they are given a basic program to start with, and the researchers created an arbitrary point system that rewards points for “positive behavior” and takes away points for “negative behavior”. The best scoring robots have their program copied to all of the other robots (the new generation) with some slight randomization. This successfully generates a new generation with heritable traits refined by a selective pressure. In short, Darwinism. It should be explicitly stated that after the initial program, there is no intelligent editing of the code. It is changed and improved only through combination, mutation, and selective pruning.
In this particular experiment, there were 100 groups of 10 robots. The robots were equipped with an Omni-directional light sensor, ground sensors, and a light ring. 11 neurons were attached to the sensors, and 3 more attached to each tank tread and the light. The ‘neurons’ were interconnected by 33 ‘synapses’. The ‘genetic code’ that determines the robot personality is an list of 8bit numbers, one for every ‘synapse’, that regulates the strength of the connection between the two joined ‘neurons, creating a variable program. The top 200 scoring robots from each trial would have their programming combined, randomized, and reintroduced as the next generation.
Their goal was to find a ‘food source’ while avoiding ‘poison’. Two objects with differently colored rings of light represented the food and the poison. Robots were given points for spending time up against the food, and given negative points for time spent in proximity of the poison. After every trial, points were added up and the best robots had their codes mixed together, changed randomly, and then given to the next generation. Simulating the survival of positive traits being combined and mutated with every generation. A link to a short video of the process is provided below.
The explanation seems dry at first, but the results are astounding. While a consciousness isn’t possessed, recognizably life-like behaviors presented themselves. Initially the robots were programmed to follow the light emitted by others, and to light their blue light when they found the food source. After only 9 generations, the robots became very proficient at detecting the light and moving towards it, making the light a very important and useful tool to effectively find the food. Room around the food source is limited however. Robots became so proficient, that initial food finders would sometimes be pushed out of proximity by attracted robots, giving them less points, and thereby cleansing them from the gene pool. By the 50th generation, robots were turning on their lights outside of the area of the food more often than inside, and the light conversely became a poorer source of information. Many robots learned to keep their light off entirely; a complete change from the original programming. The robots were deceiving one-another!
This remarkably animalistic-adaptation arose from random traits with selective pressure, demonstrating Darwinism in miniature. As one would expect in a natural environment, the robots further adapted to the deception, learning to ignore the misleading lights more and more. This lead to less selective pressure against shining one’s light near the food, because everyone else would ignore the signal. As a result, about 60% of the robots in the 500th generation lit their lights when in proximity of the food. This yielded no real benefit, but survived because it was more or less benign. In short, genetic drift. With the robots now relatively proficient at ignoring each other and finding food, the selective pressure decreased significantly. Much like with humans, where our intelligence allows us to overcome most obstacles using tools, Darwinism is made less effective at improving the populace because a minimum ability of survival is present in every robot. Therefore diversity found within the population is less due to beneficial traits, and more due to random genetic mutation that isn’t selected against.
I found this research to be interesting because I enjoy seeing scientific principles at work. Darwinism, one of the most well established and unfortunately contested systems, is so simple that it can be applied to almost any quality-dependant system. In other experiments by this lab, predator-and-prey behavior evolved to a very animalistic pattern, including lying in wait. In another, a maze robot initially so defective that it actively sought out walls, was able to navigate a closed space as well as a mouse. This method not only provides an interesting demonstration, but opens the door for practical applications. This non-cognizant learning could be applied to AI programs to work as a behavioral learning. The problem with Darwinism is that it takes generations to function, which is decades for most organisms, but only seconds for computer programs. Difficult programing problems could be solved by trial-and-error, creating novel new coding methods, and optimizing systems already at work. For someone who is split between a love of computers and biology, and who chose biomedical engineering as a compromise, this is very exciting.
Sources:
Laboratory of Intelligent Systems of the University of Laussane websie.
"Not Exactly Rocket Science" blog: http://scienceblogs.com/notrocketscience/2009/08/robots_evolve_to_deceive_one_another.php?utm_source=selectfeed&utm_medium=rss
Link to report on the study: http://infoscience.epfl.ch/record/139388/files/PNAS-2009-Mitri-0903152106.pdf?version=1
Link to the company website: http://lis.epfl.ch/
Link to video: http://www.youtube.com/watch?v=-M5oc4cBtCg