Vol. 22 · No. 6

This article examines Marvin Minsky's Society of Mind theory of human cognition. We describe some of the history behind the theory, review several of the specific mechanisms and representations that Minsky proposes, and consider related developments in artificial intelligence since the theory's publication.

Mirror neurons are a relatively recent discovery; it has been conjectured that these neurons play an important role in imitation learning and other cognitive phenomena. We will study a possible place and role of mirror neurons in the neural architecture of embodied cognitive agents. We will formulate and investigate the hypothesis that mirror neurons serve as a mechanism which coordinates the multimodal (i.e., motor, perceptional and proprioceptive) information and completes it so that the agent remains always situated even when parts of the multimodal information are missing. We show that such a hypothesis forms a basis on which plausible explanation of the development of a host of mental abilities could be founded. These abilities range from imitation learning, communication via a sign language up to the dawn of thinking. Our results build a bridge between the theory of embodied cognition and mirror neurons; they also justify the hopes related to the discovery of mirror neurons.

In a persistent multi-agent system, it should be possible for new agents to benefit from the accumulated learning of more experienced agents. Parallel reasoning can be applied to the case of newborn animals, and thus the biological literature on social learning may aid in the construction of effective multi-agent systems. Biologists have looked at both the functions of social learning and the mechanisms that enable it. Many researchers have focused on the cognitively complex mechanism of imitation; we will also consider a range of simpler mechanisms that could more easily be implemented in robotic or software agents. Research in artificial life shows that complex global phenomena can arise from simple local rules. Similarly, complex information sharing at the system level may result from quite simple individual learning rules. We demonstrate in simulation that simple mechanisms can outperform imitation in a multi-agent system, and that the effectiveness of any social learning strategy will depend on the agents' environment. Our simple mechanisms have obvious advantages in terms of robustness and design costs.

It is considered a difficult task to design a controller from scratch for a robot in a dynamic environment. Evolutionary and genetic algorithms are frequently used to find a solution with desired properties. The evolution is supposed to run on one robot or agent processor. In this article we explore the possibility of dividing the genome among several robots. Robots exchange among each other successful controllers in the form of chromosomes that code for the weight vector in a neural controller. Thus; the evolution can be faster because it runs in a parallel manner over the whole robotic society, not in one robot. Second interesting point is the exchange of experience among robots. Robots can send each other parts of the controllers that they evolved on their own. They can learn behavioral strategies that other robots developed. We describe experiments done with small Khepera robots in the domain of a benchmark test for evolutionary robotics. We compare the behavior generated by one robot based on its individual evolution with the robot that profits from sharing the full set of chromosomes with the other robot.

Agents in a multi-agent system usually use middle-agents to locate service providers. Since one central middle-agent represents a single point of failure and communication bottleneck in the system, therefore a structure of middle-agents is used to overcome these issues. We designed and implemented a structure of middle-agents called dynamic hierarchical teams that has user-defined level of fault-tolerance and is moreover fixed scalable. We prove that the structure that has teams of size łambda has vertex and edge connectivity equal to łambda, i.e., the structure stays connected despite łambda-1 failures of middle-agents or łambda-1 communication channels. We focus on social knowledge management describing several methods that can be used for social knowledge propagation and search in this structure. We also test the fault-tolerance of this structure in practical experiments.

We propose a way to define (and, in a certain extent, even to measure) the phenomenon of emergence which appears in a complex system of interacting agents whose global behaviour can be described by a language and whose components (agents) can also be associated with grammars and languages. The basic idea is to identify the ``linear composition of behaviours'' with ``closure under basic operations'', such as the AFL (Abstract Families of Languages) operations, which are standard in the theory of formal languages.