Gabriel J. Ferrer

The effective field of view paradigm: adding representation to a reactive system

Abstract Dynamic environments pose multiple problems for autonomous agents. Traditionally, agents were designed to develop an internal model of the world and use it to create plans to achieve their goals. Such a strategy fails when the actual environment differs from the stored representation. This realization led to reactive agents, which stored no model of the world. Today, most researchers would agree that representation is not intrinsically bad, but the cost of keeping an arbitrary amount of representation up-to-date in a dynamic environment outweighs its benefits. This paper presents a simple task-dependent representation system, called the effective field of view, which can be used to augment a reactive agent and improve its competence at a variety of tasks. This system has been embodied in two situated autonomous agents. With the first, the amount of representation necessary for its tasks is quantified, as well as the performance gains. With the second, it is shown how the effective field of view can be used in a hierarchical layered agent architecture. This paper shows that the effective field of view can dramatically increase agent performance.

Using genetic programming to evolve board evaluation functions

Employs the genetic programming paradigm to enable a computer to learn to play strategies for the ancient Egyptian boardgame Senet by evolving board evaluation functions. Formulating the problem in terms of board evaluation functions made it feasible to evaluate the fitness of game playing strategies by using tournament-style fitness evaluation. The game has elements of both strategy and chance. Our approach learns strategies which enable the computer to play consistently at a reasonably skillful level.

Layered mode selection logic control for border security

Challenges in border security may be resolved through a team of autonomous mobile robots configured as a flexible sensor array. The robots will have a prearranged formation along a section of a border, and each robot will attempt to maintain a uniform distance with its nearest neighbors. The robots will carry sensor packages which can detect a signature that is representative of a human (for instance, a thermal signature). When a robot detects an intruder, it will move away such that it attempts to maintain a constant distance from the intruder and move away from the border (i.e. into its home territory). As the robot moves away from the border, its neighbors will move away from the border to maintain a uniform distance with the moving robot and with their fixed neighbors. The pattern of motion in the team of robots can be identified, either algorithmically by a computer or by a human monitor of a display. Unique patterns are indicative of animal movement, human movement, and mass human movement. To realize such a scheme, a new control architecture must be developed. This architecture must be fault tolerant to sensor and manipulator failures, scalable in number of agents, and adaptable to different robotic base platforms (for instance, a UGV may be appropriate at the southern border and a UAV may be appropriate at the northern border). The Central Arkansas Robotics Consortium has developed an architecture, called Layered Mode Selection Logic (LMSL), which addresses all of these concerns. The overall LMSL scheme as applied to a multi-agent flexible sensor array is described in this paper.

A linear time transform for probability aware planning

Presents a transform that enables traditional shortest-feasible-plan planners to reason about uncertain operators and produce plans which have higher probabilities of success. This transform converts a probability-aware domain description into a STRIPS-style description, where the probability of success is expressed by the plan length. Using this transformed description, a plan can be generated by a traditional planner. The transform is shown to be at worst linear in the size of the input, and allows the planning system to trade-off accuracy against runtime as an anytime computation.

Anytime replanning using locality in agent architectures

Anytime planners seek to allow the system to exchange execution time for solution quality. We propose a new anytime planner which will devise new plans to accommodate plan failures. The first new plans to be considered will be relatively local to the plan that failed. Further new plans will be less and less local. We intend the concept of locality to relate the search space to be explored for a new plan to the degree to which a new plan disrupts the not-as-yet executed part of the running plan.