Optimizing Interaction Potentials for Multi-Agent Surveillance POC: Dr. William Spears Computer Science Department Dept 3315 1000 E. University Avenue Laramie, WY 82071 wspears arobase cs.uwyo.edu		Team Members: Faculty: Dr. William M. Spears, Computer Science Department Dr. Diana F. Spears, Computer Science Department Graduate Students: Wesley Kerr, Computer Science Department Suranga Hettiarachchi, Computer Science Department Dimitri Zarzhitsky, Computer Science Department

Project Description

This research examines several potential energy models within the physicomimetics (also known as artificial physics, or AP) framework used to control aerial surveillance assets (such as UAVs) tasked with detecting ground-based targets (such as tanks and transport vehicles). The assets are assumed to cruise at a constant altitude, equipped with target and terrain sensors. The environment consists of open areas where target detection is possible, and areas hidden by foliage, where target sensors are ineffective. The aim of the project is to develop the optimum surveillance strategy, which maximizes area coverage and target detection.

To achieve our objective, we constructed a simulation tool (see Fig. 1) that facilitates experiments with several realistic conditions, such as sensor noise, agent failure, and different adversary behaviors. A genetic algorithm is employed for the purpose of evaluating system performance across large sets of possible parameters, and resilient strategies with high target detection rates are evolved. The schematic for the basic system architecture is given in Fig. 2 below.

Short Summary

• Interactions between assets can outperform methods where no interaction occurs.
• Optimizing potentials is a promising method for creating good interactions.
• Such interactions are even more useful when tanks move and adopt their own intelligent strategies.
• These potentials show strong evidence of generality, despite not being trained for that. With explicit training for generality, performance will improve further.

Methodology and Experiments

• Simulate 10 random environments with variable foliage coverage. In each scenario, assets move according to the potential function, while trying to detect 100 tanks, which may remain stationary, or may move according to several general trajectories.
• Study the impact of sensor noise, which we model as a single probability of target detection.
• Use a genetic algorithm to determine the optimum potential for environment #1.
• Test all of asset strategies in each environment; potential determined by the GA for environment #1 is used for all 10 environments.

Publications

Acknowledgment

This project was funded by the Defense Advanced Research Projects Agency (DARPA).