My interest lies in understanding and imitating the processes that allow animals and humans to take flexible decisions in a complex and difficult environment, and enable them to adapt to different conditions gracefully. Can we do so incorporating learning ability, without compromising generalization ability, without hand-coding all necessary rules into a system? Are there general principles underlying intelligent information processing in living beings which we can exploit without having to resort to specialized solutions that vary from task to task?
For this purpose, my work employs methods from Artificial Life, and especially Information Theory, and apply them to Sensor Evolution, Collective and Complex Systems.
Our recent contributions lie in a principled formalism for the information-theoretic treatment of the perception-action loop, based on a recently developed notion of information flow, and a information-theoretic bottom-up approach for the reconstruction of a world model for a real robot. In addition, we introduced "empowerment" as a principled information-theoretic "universal utility", creating a value system based entirely on the embodiment of an agent; this provides an indication how living systems are able to systematically impose structure onto their high-dimensional state space. Furthermore, we have begun to understand how cognition and behaviour are constrained in informational terms which allows us to consider informational conservation laws for cognitive behaviour. This opens the route for systematic studies of cognitive architectures which would be pertinent to given tasks and agent embodiments.
In other words, we study information bookkeeping in a cognitive agent. This research promises a route to address a set of major future challenges: how to cope with widely distributed, but ill-defined robotic designs (e.g. cheaply built, or 3D printed robots), and how to do that through low-capacity, but ubiquituous information processing devices:
- it allows us to study the best way to organize information processing for a given task by a given agent;
- it might be able to tell us where an informational compromise is the least costly;
- and, the methods indicate that it might even be possible to predict what the best "body" would be to solve particular tasks - with today's flexibility of 3D printing, these methods might provide us with a guiding map in the proliferating jungle of potential robotic designs;
- finally, once the robot is running, the "empowerment" intrinsic motivation model provides it with its initial cognitive "yolk" to start operating in a new environment to some effect, without necessarily requiring a hand designed behaviour strategy by a topic specialist.
In summary, the idea is to study the information flow in a cognitive agent which will improve our understanding of the constraints of natural cognition and how to endow agents and robots with biologically plausible, effective, and flexible cognitive architectures.