THE ADAPTIVENESS OF INTELLIGENT AGENTS

Classical Artificial Intelligence (AI) assumed human intelligence as its paradigm, and paid almost no attention to "lower level" animal intelligence. Since the beginning of the eighties, however, an alternative view has gradually gained favor. The continuity between animal and human intelligence has been appreciated, and substantial research efforts have been devoted to biologically inspired models of intelligence, like neural networks and evolutionary computation. Moreover, physically embedded agents, which must interact via noisy sensors and actuators with partially unpredictable environments, have become one of the favorite testbeds for such models.

One of the main aspects of intelligence, under this broader acceptation, is adaptiveness, that is, the ability of an agent to adapt to environmental conditions in order to survive and carry out its task. Adaptiveness is not sufficient for intelligence, and in fact it is a property of all forms of life, even of the ones that we would not dub as intelligent (for example, plants and bacteria); however, it is certainly a necessary condition for intelligent behavior.

In the realm of biological systems, adaptiveness is an essential component of an organism's "fitness", that is, of the ability to survive and spread its genetic make-up. Analogously, in the realm of the artificial, adaptiveness is going to be one of the key components of agent quality.

From the point of view of agent engineering, the problem of adaptiveness manifests itself at least at two different levels. At the lower level we have the problem of designing an adaptive sensorimotor system, that is, one in which both sensors and actuators can adapt in real time to environmental changes (think for example of a visual input system that is insensitive to lighting conditions); at the higher level we face the problem of implementing adaptive behavior. This makes it possible to cope with new environments (usually, an agent is tested in its native environment, but it should operate in the customer's), and to adjust to dynamically changing environments.

The implementation of truly adaptive physical agents is still a matter of study. However, we believe that agent technology is now mature enough to directly face this problem. The two main roads to the implementation of adaptive behavior seem to be:

• using biologically inspired hardware and software architectures (based on neural networks, learning classifier systems, fuzzy models, etc.) to implement agents' controllers;
• adopting machine learning methods, like supervised learning, reinforcement learning and evolutionary computation, to allow agents to acquire and optimize their behavior through experience.

These two aspects are both important to obtain adaptivity. Biologically inspired architectures are robust and can be modified without strong discontinuity in behavior. Machine learning techniques give agents the capacity to adapt their behavior without human intervention.

In the last four years, at the Artificial Intelligence and Robotics Project of the Politecnico di Milano we have carried out substantial experimentation on the use of learning classifier systems, both classical and fuzzy, and of neuro-fuzzy supervised learning for the control of intelligent physical agents [1,2,3,4]. Our aim has been to realize physical agents that show some degree of adaptiveness while carrying out their tasks. The results of our experiments are encouraging, and we are now considering a set of specific applications for mobile robots (in particular, mail distribution and cleaning in an office environment).

[1] Bonarini, A., (1996), Delayed reinforcement, fuzzy Q-learning and fuzzy logic controllers. In Herrera, F., and Verdegay, J.L. (eds.), Genetic Algorithms and Soft Computing, Physica-Verlag, Berlin, D, 447-466.

[2] Colombetti, M., Dorigo, M., and Borghi, G. (1996). Behavior Analysis and Training: A methodology for behavior engineering. IEEE Transactions on Systems, Man, and Cybernetics - Part B, 26 (3), 365-380.

[3] Dorigo, M. (1995). ALECSYS and the AutonoMouse: Learning to control a real robot by distributed classifier systems. Machine Learning, 19 (3), 209-240.

[4] Dorigo, M., and Colombetti, M. (1994). Robot shaping: Developing autonomous agents through learning. Artificial Intelligence, 71 (2), 321-370.

* Artificial Intelligence and Robotics Project, Dipartimento di Elettronica e Informazione, Politecnico di Milano

^ IRIDIA (Institut de Recherches Interdisciplinaires et de Développements en Intelligence Artificielle), Université Libre de Bruxelles