Page 312 - Introduction to AI Robotics
P. 312

295
                                      8.1 Overview
                                        It is not clear when communication is needed between agents, and what to say.
                                         Many animals operate in flocks, maintaining formation without explicit
                                         communication (e.g., songs in birds, signals like a deer raising its tail to
                                         display white, speaking). Formation control is often done simply by per-
                                         ceiving the proximity to or actions of other agents; for example, school-
                                         ing fish try to remain equally close to fish on either side. But robots and
                                         modern telecommunications technology make it possible for all agents
                                         in a team to literally know whatever is in the mind of the other robots,
                                         though at a computational and hardware cost. How can this unparalleled
                                         ability be exploited? What happens if the telecommunications link goes
                                         bad? Cell phones aren’t 100% reliable, even though there is tremendous
                                         consumer pressure on cell phones, so it is safe to assume that robot com-
                                         munications will be less reliable. Is there a language for multi-agents that
                                         can abstract the important information and minimize explicit communi-
                                         cation?
                                        The “right” level of individuality and autonomy is usually not obvious in a prob-
                                         lem domain. Agents with a high degree of individual autonomy may cre-
                                         ate more interference with the group goals, even to the point of seeming
                                         “autistic.” 113  But agents with more autonomy may be better able to deal
                                         with the open world.

                                        The first question in the above list essentially asks what are the architectures
                                      for multi-agents? The answer to that question at this time is unclear. Individ-
                                      ual members of multi-agent teams are usually programmed with behaviors,
                                      following either the Reactive (Ch. 4) or Hybrid Deliberative/Reactive (Ch. 7)
                                      paradigms. Recall that under the Reactive Paradigm, the multiple behaviors
                                      acting concurrently in a robot led to an emergent behavior. For example, a ro-
                                      bot might respond to a set of obstacles in a way not explicitly programmed
                                      in. Likewise in multi-agents, the concurrent but independent actions of each
                     EMERGENT SOCIAL  robot leads to an emergent social behavior. The group behavior can be different
                            BEHAVIOR  from the individual behavior, emulating “group dynamics” or possibly “mob
                                      psychology.” As will be seen in this chapter, fairly complex team actions such
                                      as flocking or forming a line to go through a door emerge naturally from re-
                                      active robots with little or no communication between each other. But as
                                      with emergent behavior in individual robots, emergent social behavior is of-
                                      ten hard to predict. Complete architectures for designing teams of robots are
                                      still under development; Lynne Parker’s ALLIANCE architecture 114  is possi-
                                      bly the most comprehensive system to date. The whole field of multi-agents
                                      is so new that there is no consensus on what are the important dimensions,
   307   308   309   310   311   312   313   314   315   316   317