Introduction:

Multi-agent robot systems are clearly the next challenge in the field of robotics. There are many good reasons for this. Multi-agent robot systems promise to handle complexity while providing for higher robustness, reliability, and performance, and all this while promising to lower cost. Nature itself is an abundance of examples where cooperation/competition among members of a multi-agent system allows for handling a complex task. Societies of bees and ants are instances of successful cooperation among a large number of entities. Human hand is another example of several cooperating fingers handling an ill-defined object. Also, for migrating birds, the formation flying provides for decreased overall drag, and hence, increased range. An example of competing societies of species is natural evolution with its well known benchmark theme, the “survival of the fittest.” Besides multi-agent robotics, applications of multiagent systems include internet traffic routers for improved transmission speed of relevant information and coordination of a constellation of satellites for increased efficiency in interferometeric synthetic aperture radar.

Figure 1. Multi-agent robotics is also a multi-faceted field.

Theoretically, multi-agent systems are a gold mine of interesting theoretical problems. Multi-agent robotics is not a homogeneous field of research, but rather is a merger of various fields such as distributed control, artificial intelligence (AI), sensors and instrumentation, nanotechnology and miro-electonics-mechanical-systems (MEMS), and software coding challenges, see Figure 1. Among above general issues, our group is most concerned with developing an appropriate framework for modeling cooperation, learning and intelligence. For this, we are considering fuzzy logic as a framework of knowledge representation, stochastic learning automata and genetic algorithms as means of learning, and discrete even systems as framework of modeling. Furthermore, the questions that we try to answer are

a.       How can the above various methodologies be unified, and

b.      How we can ascertain their mutual superiority.

DEVS-SLA (Contributed by Yan and Shahb)

The DEVS ( Discrete Even System Specification) environment is distributed modeling environment, which was cereated to provide a robust and nonspecific environment for modeling and simulation projects.

Control theory inclduing optimal control theory requires perfect information or a priori information of the systems to be controlled. In most practical situations, where uncertainty is paramount, the use of learning control algorithms becomes  a necessary alternative. Learing control is of particular interest in distributed robotics. In most cases the robots are sent into an invironment of which little knowledge is availabe, i.e., contaminated depository. The robots are then required to learn their environment and perform their tasks with minimal error.

              An important tool in learning control is SLA (Stochastic Leanring Control). The basic operation of SLA is as follows. At any given time, an action is performed by SLA based on internal states. Due to that action, the environment (also referred to as a teacher) responds with a value between 0 and 1. A value of zero corresoponds to full reward and a value of one corresponds to full penalty, Based on this feedback from the environment, SLA updates the probabilities of choosing a certain state. This process is repeated until the average penalty is minimized.