Last updated November 9, 1995

The research interests of Marco Dorigo are:

Robots Shaping and Behavior Engineering

The long-term research goal is to develop an integrated methodology for the development of autonomous agents that interact with a physical environment. Specific research topics are the specification of target behaviors, the design of sensorimotor interfaces, the design of the agentÍs control architecture, the use of machine learning algorithm to develop control programs strongly coupled with the environment, the design of domain-dependent training strategies, and the evaluation of performance. The research methodology is strongly experimental; experiments are run both in simulated environments and in the physical world, and exploit Alecsys, a development environment for learning agents designed by Marco Dorigo. The results achieved so far concern: the use of distributed architectures for agents capable of complex behavior patterns, and the development of nonreactive (i.e., dynamic) agents. See related publications

Ant Colony Optimization

Theoretical Analysis of Genetic Algorithms

We have proved a theorem about a lower bound on the quantity of information processed by a classical genetic algorithm. We are now trying to extend the proof to the more general class of evolutionary algorithms (the genetic algorithm is a particular instance of evolutionary algorithm). See related publications

Reinforcement Learning

We have studied an extension to the bucket brigade. We have shown the formal equivalence of Q-learning with the bucket brigade algorithm. We have investigated a series of improvements to the standard Q-learning algorithm. Experiments have been run on both grid world and Animat-like problems. We have also proposed a reinforcement learning algorithm for probabilistic boolean networks. See related publications

Evolutionary Computation in Optimization

In this research the main interest is in the use of evolutionary algorithms to solve very difficult (NP-complete) problems. As a test problem we first used the timetable problem and we developed a program that was applied to solve a real high school timetable. We also developed a program based on evolutionary algorithms to schedule the operations of a robotic system. See related publications

Morphogenesys

The main goal of this research is to develop simulated robotic agents by the interplay of evolutionary and learning techniques. We start from basic very low-level building blocks, cells, which contain genetic information. A morphogenetic process letŐs cell duplicate and differentiate subject to environmental constraints. Organisms so created can start, after the development phase, to reproduce. In this way fittest, that is those individual who live longer, will have more offspring. During duplication new generated individuals undergo genetic changes like mutation and crossover. Learning capabilities are not designed into the system, but emerge as a consequence of their adaptive value. See related publications

Algorithms Parallelization

We have studied the properties of some search algorithms (like A*) and we implemented a parallel version of A* on a 18-transputers system. We have also, during a period spent at the technical university of Munich, designed a generator of simulators of neural networks on a smaller transputer system (4-transputers board). See related publications

IRIDIA, Université Libre de Bruxelles
Avenue Franklin Roosevelt 50
CP 194/6
1050 Brussels - Belgium
Tel: + 32 2 6503169
Fax: +32 2 6502715
email: mdorigo@ulb.ac.be