The Attractor Agent Network (AAN) is my multi-agent system (MAS) architecture that uses a Hopfield network to activate and coordinate a set of agents. The aim of the architecture is to learn scheduling policies and to recall them online in a context aware manner, so that the right subset of agents is activated in the right span of time to make them cooperate and fulfill some specific goal. AAN has an architectural nature: the users have only to establish a particular implementation of individual agents and a function that represents their ability to achieve their goals individually. Once this is done AAN will automatically bring the system to an effective collective state.
This is achieved by connecting the agents with a Hopfield network that distributes a bounded amount of computational resources among them. The constraint of a limited amount of computational resources introduces competition among agents, which allows the architecture to automatically select the best subset of agents to be active at the same time.
A Hopfield network is used in order to benefit from the well defined properties coming from its definition in statistical physics. This will lead to a dynamical characterization of the agents' allocation policies called "agents' attractors". The usage of Hopfield networks makes available to MAS a sophisticated mathematical language, composed of formal concepts such as probability distribution of activation, attractors, energy function, correlation function and other macroscopic variables that summarize the collective behavior of agents. These can be used to analyze, predict and tune the MAS behavior.
A complete description of the AAN architecture can be obtained by reading my PhD dissertation: "the modularity of attention from an artificial intelligence perspective".
The Attractor Predictor Network is an implementation of the AAN architecture that distributes computational resources among a set of predictors representing sensorimotor contingencies. The architecture controls an artificial eye in an active vision framework to solve a perceptual decision making task: the categorization of stick animals.
Here the ability of the AAN architecture to automatically coordinate a set of agents is analyzed in close link with statistical physics. We show how the free energy can be used as a measure of agents coordination. In addition, an online algorithm is proposed to automatically learn agents' attractor. Finally, we report the results of a distributed constraint optimization experiment related with a board game.
The Attractor Expert Network (AEN) is an implementation of the AAN architecture that connects a set of feed forward neural networks called "experts". The AEN extends the mixture of experts model (MOE) adding a set of connections between the experts to store their contextual relationships. This makes the architecture able to exploit cooperation issues among the experts and not only competitive ones as in the classical MOE model. In addition, the MOE's gating network is used to map the input to the initial activities' configuration and as a procedure to specialize the experts. Finally, we validate our proposal to solve a classification task from machine learning.
Automated reasoning is a computationally hard problem when the set of formulae is too large. This happens because classical models are not able to use semantic informations to select only the most suitable set of formulae necessary to prove. Using the Attractor Prover Network an automated reasoning task can be distributed among several small theorem provers, storing such kind of informations in the agents' attractors to obtain a robust and efficient prover.