Algorithm Comparison for Moving Target Search CMPUT 651 Midterm Report

For moving target search algorithms, we are considering the case of heuristic search where the goal may change during the course of the search. One motivating application lies with navigation of an autonomous police vehicle chasing a villain [5] in a possibly initially unknown environment under real-time constraints. In this scenario, we would like to assure that the autonomous vehicle can quickly reach the target, preferably faster than human driving. We would also like such vehicle to learn the environment so that the vehicle performs better (catching the villain faster) in the previously explored area. This is vital when there is a cost associated with the moves the vehicle made. Similarly, moving target search algorithms can be applied to new generation of real-time strategy games, where the agent is required to cope with the initially unknown maps via exploration and learning during the game [3]. As a result, these games will greatly limit the applicability of complete search algorithms and pre-computation techniques [1]. Considering one scenario in Starcraft, a group of Protoss units have 60 minutes to destroy all Terran units. We assume that each group is placed on an initially unknown map and can only learn the map via exploration. The units in the same group share global information of the world. When a Dragoon (Protoss unit) sees a Marine (Terran unit) in its sight and knows that it is stronger than the Marine. As a result, the Dragoon begins to chase the Marine, who decides to run away. The Dragoon wants to kill the Marine with the minimum move possible before the Marine gets to Terran base. It also wants to learn the world in that area it has visited so that all Protoss units can reach Terran units quicker for later moves, which is very crucial given limited time to succeed. In both cases, we do not need the agent and the target to be at the same location to terminate the search. We will, instead, terminate the search when the agent is close enough to the target. For example, Dragoon units in Starcraft, which act as