The modularity of attention from an artificial intelligence perspective

The development of agents with bounded rationality is still an important challenge of artificial intelligence. Indeed, when we are facing problems with a large number of states, if we do not reason about the computational resources of our agent, it is easy to encounter exponential complexities or state explosions. One way to solve this problem is taking inspiration from the study of attention in cognitive science. Attention can be considered as a filter in information processing that focuses only on relevant information, leaving out possibly all the useless computations. So to be focused only on a relevant subsets of the state space of a problem can be the solution to increase the quality of our algorithms. Accordingly it is important to understand how to represent relevance and being able to compute automatically what is relevant in a given situation. In this thesis we link the study of attention with multi agent systems (MAS). The introduced methodology starts finding a way to partition an input problem. Then it specializes the agents of a MAS on the partition’s subsets. Finally it finds a policy to switch the agents on or off according to the current context. This will be done allocating computational resources to the agents during the task execution. Accordingly we will introduce two context aware methodologies based on Hopfield networks and dynamical neural fields to learn, store and recall online resource-allocation policies. This will lead to a dynamical characterization of such policies (i.e., attractors). Finally the systems will be evaluated in distributed constraint optimization, classification and categorization tasks, underlining, when was possible, the cognitive plausibility of our proposals. Indeed the categorization task will be held in an active vision framework linking, also experimentally, our proposals with the study of attention in animal and human vision.