If you have any problems related to the accessibility of any content (or if you want to request that a specific publication be accessible), please contact us at email@example.com.
Genetic Algorithms Optimized Potential Fields For Decentralized Group Tasking
AdvisorLouis, Sushil J.
AltmetricsView Usage Statistics
Maneuvering autonomous agents to accomplish complex tasks is a difficult and typically NP-hard optimization problem with many real-world applications. In this thesis, we use potential fields based on task and agent properties to control the movement of groups of agents and use a genetic algorithm (GA) to optimize potential field parameter values to lead to complex task achieving behaviors. More specifically, we control autonomous unmanned aerial vehicles (UAVs) in search and rescue scenarios to find and help people in need, in wildfire coverage scenarios to monitor a wildfire's perimeter, and game agents in real-time strategy (RTS) games to win skirmishes. In all three applications, potential fields control agent movement, genetic algorithms optimize potential field parameters, and a simulation evaluates task performance to guide genetic optimization. Experimental results show that our potential field representation and problem formulation works well across the three problems. We used UAVs as flying access points and controlled their movement using genetic algorithms optimized potential fields to generate wireless networks. These ad-hoc wireless networks outperformed the current state of the art ad-hoc network deployment algorithm. The same representation with a different set of potential fields was used for successful deployment of UAVs to track the spread of wildfire boundaries and results show that with enough UAVs, complete fire boundary coverage was achieved. Lastly, we used two different RTS game platforms to evolve tactics for a team of heterogeneous game agents by formulating the problem as a multi objective optimization problem. Again using potential fields, a genetic algorithm evolved a diverse set of high quality skirmish tactics ranging from attacking to fleeing against test opponents. Results show that with aggressive attacking tactics, a team of friendly agents was able to eliminate the majority of opponents but suffered significant damage. On the other hand, fleeing tactics resulted in less damage to friendlies but also inflicted less damage to opponents. We also observed the emergence of cooperation between friendly game agents. These results indicate that genetic algorithms optimized potential fields are a viable approach to decentralized group tasking.