Evolution of Behavior
I am working on questions regarding the evolution of behavior and intelligence. I would like to understand why organisms behave in certain ways, and how they accomplish their tasks. We find many processes and behaviors that are unexplained, and cooperation is only one of them. I am also interested in swarming, hierarchies, multi level selection, and cognitive aspects. Instead of trying to piece together the possible evolutionary past of organism comparing present forms or using fossils, I focus on the evolutionary mechanism itself. I want to know which evolutionary parameters give rise to certain behaviors, or which environments are necessary to see specific cognitive abilities to emerge. There is but one problem – experiments in evolution take time, and is extraordinary inconvenient when it comes to higher order animals, so I take the only viable route: digital evolution. I developed a computational framework that allows me to evolve Markov Brains , which is a similar to neuronal networks, it is just a little better to evolve, and the functionality of these networks is much easier to understand. Check out the video about “hyenas” cooperating, or the 3D walkers, which are some of the applications Markov Brains are useful for:
The above video shows you virtual hyenas (red triangles) trying to move lions (yellow blocks) away from a zebra (black and white stripes). The lions are not animate here, you can think of them as boxes. What is crucial to the cooperation aspect of this work is that idea that lions can only be moved if two hyenas push at the same time. We evolved these agents over many generation using Markov Brains. What you see is a hyena on the left, first benefiting from the help of another hyena, but then later moving over to the right side of the zebra helping the next hyena push away a
lion”. A behavior only observed if we perform group level selection.
The video above shows virtual critters in a 3D physics environment also controlled by Markov Brains. Here we study if we can evolve these brains to not only work in
abstract environments, but if they would be able to perform in complex environments. First you see the second generation and later in the video you see the population after 100 generations. The task is to get from the midline as far away as possible.
When evolving organisms in complex environments we always make the same experience: They don’t behave like we expected! Either organisms find simple exploits, but often the higher order behavior like cooperation or division of labor, does not happen. The reason is, that the fitness landscapes I design, also have to satisfy game theoretic considerations. In other words: you see cooperation evolve only if cooperation is the best possible strategy. That is why I spent a good deal of time on evolutionary game theory. While this field has been studied for more than 30 years, something interesting can be found: Most people in EGT simply ignore that they deal with evolving systems … and assume the solutions can be derived from pure mathematical equations. This is true for every simple game involving two players, and up to three choices; games so simple that we wouldn’t eve expect to find them in nature. We find a lot of interesting things, once we actually evolve the strategies: Cooperation is dependent on the agents ability to predict the world, and how reliably agents can communicate, that winning isn’t everything, and that punishment facilitates cooperation but once a population cooperates you don’t need to punish, you just need the option to punish.
Please also see my comments on evolving intelligence.