Automatic Algorithm Selection in Videogame Pathfinding
Year: 2018
Type: dissertation
Abstract: Videogames often use artificial intelligence to control characters in the game world. In doing so, videogames require one or more agents to navigate from their current location to some desired goal location without collisions. We explore improving algorithm performance in both single and multi-agent pathfinding (MAPF) through the use of algorithm portfolios and machine learning algorithm selection. In the single agent setting we explore selecting from different versions of a parametrized real-time heuristic search algorithm focused on minimizing the distance the agent travels. In the MAPF setting we select from three contemporary MAPF algorithms, Windowed Hierarchical Co-operative A*, Flow Annotated Replanning, and Bounded Multi-agent A* focused on maximizing the number of agents which reach their goal. Algorithms will often have specific problem instance where their performance either excels or deteriorates, enabling another algorithm to out perform it even if this other algorithm is worse on average. Therefore, in this thesis, we investigate the use of deep learning to automatically select the best algorithm from a portfolio of algorithms for given problem instances. Empirical results show that an off-the-shelf convolutional neural network is able to improve performance over any single algorithm from our portfolio.
DOI: 10.7939/r39s1m209
Author: Sigurdson, Devon
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2952016324