The main idea behind my research is to use evolution as the means to get artificial intelligence. AI research is about 70 years old, and we made very good progress on designing very capable algorithms that mimic intelligent features of human like speech recognition, visual perception, planning moves in games, and inference from logical systems and databases. However the very intuitive nature of some basic cognitive abilities is still widely missing: telling friend from foe, dealing with decaying hardware, navigating in unfamiliar terrain, reasoning, thinking in abstractions …
In my opinion AI research of the last 70 years is not pursuing the right goal and is using a design and engineering approach. However we know of one mechanism that is and was very capable of turning none cognitive precursor material in sentient, conscious, and sane individuals: evolution!
I am not the first one to propose to evolve AI instead of designing it. You might find my publications to be very wide spread including many different fields, but you will see that most of what I do is actually necessary in succeeding in my goal.
What is intelligence?
I am not giving you a single and simple answer, but I think it has something to do with the ability of making the right decision. Fortunately there is already a theory about decision making, namely game theory. You can use game theory to make predictions about outcomes of decisions, take decisions of other into account, evaluate likelihoods of events, and most importantly you can extend game theory to population game theory and evolutionary game theory. If you want to know what decisions evolve in population, you need to understand evolutionary game theory, and in particular if you want to evolve decision makers you need to know how to use evolutionary game theory to your advantage. Imagine you want to evolve a watch dog. You have to make sure that it doesn’t befriend the enemy. Or if you want to evolve collaborators, you need to avoid selfish behavior. If you want to have division of labor, you need to evolve social stratification, and if you want leaders you need to know how to prevent the extinction of followers. I study game theory because without it you can’t understand or control evolution.
What to evolve?
For the longest time artificial neuronal networks (ANN) together with back propagation where the holy grail in AI. However, if you put an ANN (or any of it’s derivatives) in charge of controlling a bot or an agent you get very none natural looking behavior, you find that back propagation only works to a certain degree, and evolving an ANN is not the most straight forward thing to do. Therefor I developed a software framework that allows you to evolved hidden Markov models (HMM). The idea to use HMMs is also not new, however they are untrainable using conventional back propagation which rendered them useless. Also HMM tend to be very large in terms of computer memory and thus also unevolvable and impractical to use. I devised a system in which the HMM is decomposed into subcomponents and encoded genetically in order to make it evolvable. The result is striking. We see much more natural behavior to evolve, we end up with a cognitive architecture that can be understood, manipulated, and compared – something impossible to do with ANNs. These evolved HMM, which are called Markov Brains, are so lightweight that they can be rendered into FPGAs, ported to small micro controllers, and used as AI in computer games without burdening the game engine with too much computational overhead.
How to understand brains?
If you want to understand something you first of all need to measure it. Therefor I working on two fundamental cognitive processes: How information is integrated, and how representations are formed. And in addition, as stated above, Markov Brains are topologically interesting graphs, which need to be analyzed and compared to existing neuro-architecture. My work on graph theory resulted in a couple of interesting insights into the modular organization of graphs and the building blocks within neuro-anatomy.
Where to go from here?
When using evolution as the means to reach AI you have to start somewhere. Our computational infrastructure is not nearly as advanced to evolve human level intelligence right now, and we don’t unerstand evolution enough to design the necessary fitness landscapes. Therefor I study how fundamental cognitive processes were shaped by evolution. I study the evolution of swarming, foraging, collaboration, information integration, and spatial temporal reasoning.
AI in computer games
Your bus schedule or car chassis might have been optimized by evolutionary algorithms, but you have not interacted with any form of evolved AI. Where will you meet evolved AI first? My suggestion is computer games. They are playful environments, where the worst thing that can happen is a boring experience. Games are easy to control and their data is persistent and obvious reducing the complexity of many problems drastically. And lastly, computer games are a widespread phenomenon – large scale computational infrastructure and an enormous amount of CPU hours is spent on gaming. However most importantly it is a place where humans like spent time, providing numerous opportunities for interactions.
One of the largest obstacles in generating games is to provide content, context, and texture (or pulp, how I call it). Currently, everything you see in a game is the product of someone designing content. This is laborious and expensive. Designing content that is dynamic and adaptive is more or less impossible. Evolution is the one source of novelty in nature, it created complex ecologies and economies, and because of emergence it is very capable of creating surprise – the one thing that is hardest to craft. I don’t think I need to say more.