Simon Bovet

Simon Bovet PhD Student, Artificial Intelligence Laboratory, University of Zurich

Andreasstrasse 15
8050 Zürich
Switzerland
+41 44 635 43 42

Research

When natural intelligent systems are studied with artificial agents, control architecture and body dynamics are usually designed for specific tasks. This contrasts with natural evolution, considered to have no goal or intention. An alternative approach consists in investigating how emergent behaviors can be observed with systems whose controls are not designed for any particular task. The main goal of my research is to investigate how different behaviors can emerge from a simple control architecture consisting only of systematic Hebbian-like multi-modal coupling.

AMouse It is obvious that any intelligent system (either natural or artificial) must somehow possess the ability to maintain some information about the interaction with its environment. A notorious problem that arises when building an artificial system concerns the implementation of such a feature. What kind of information is relevant for the agent? How to "store" this information so that it can later be "retrieved" in a consistent way? Very often, designers opt at least partially for ad-hoc solutions tailored in actual fact for the task the agent has to perform.

MiniDog For instance, designers almost always provide the control system of autonomous robotic agents with a set of basic behaviors: these can be unconditioned reflexes, basic modules or layers achieving incrementally complex tasks, or even a general hierarchy in the the structure of the control architecture. It is not clear how arbitrary the assumptions made by the designer about how the external world is perceived by the situated agent actually are. Moreover, it is largely unknown to what extent the biases introduced by these design assumptions may reduce the diversity of possible emergent behaviors (i.e. behaviors emerging from the agent-environment interaction which are not explicitly programmed into the system), commonly accepted as important prerequisite for "intelligent" systems.

The main objective of my research is to demonstrate that an agent controlled by a simple architecture model consisting merely of homogeneous sensorimotor coupling with Hebbian plasticity, can have the potential of displaying several emergent and complex behaviors.

In short...

Put in simply: how is it that an agent, whose "brain" consists only of simple neurons and synapses (i.e. which is not "preprogrammed" for any specific task!), can learn to follow an object, to navigate back to its home position using strategies observed in insects, or even to "understand" the temporal relationship between an early clue and a delayed reward without any sensory memory?