Pathfinding Algorithm Optimization Via Evolution

Pathfinding is a popular computer science problem in both academic research and industrial development. The objective of pathfinding is to search for a path, often the shortest path, from one location to another on a graph. Many real world applications can be considered as pathfinding problems, including motion planning, video games, logistics, and decision making. Computer scientists have proposed different algorithms to efficiently search for the shortest path. A* search algorithm is the de facto pathfinding algorithm that uses a heuristic function to determine the best action to take based on the given information. It is the most popular pathfinding algorithm due to its simplicity and efficiency. The performance of A* is heavily dependent on the quality of the heuristic function. The heuristic function determines the search speed, accuracy, and memory consumption. Hence, designing good heuristic functions for specific domains becomes the primary research focus on pathfinding algorithm optimization. In this dissertation, we address and solve several commonly known challenges in pathfinding problems and A* algorithm. First, designing new heuristic functions is a difficult and time-consuming task, especially when they are used to solve complex problems. The task requires the user to have expert knowledge of the problem. Moreover, a single heuristic function might not be enough to digest all the provided information and return the best guidance during the search. Previous works suggest that multiple heuristics for complex problems can dramatically speed up the search. However, choosing the appropriate combination of heuristic functions is tricky. Current optimization approaches rely on hand-tuning the parameters via trial and error by engineers over many iterations. There is a need to reduce the difficulty of designing heuristic functions for search performance maximization. Our first contribution is to propose an improved A* with a self-evolving heuristic function named Evolutionary Heuristic A* (EHA*) that reduces engineering effort to design the heuristic function for A* and maximize the search performance. Our experiment results show that EHA* (i) preserves path optimality; (ii) is not limited to a particular application; (iii) speeds up the path searching process; and (iv) most importantly, dramatically reduces the difficulty for software engineers to design heuristic functions for A* search. Moreover, our work can be applied to other existing works on the performance improvement of A* search. Search, A* search suffers from poor performance on large search spaces. Although EHA* improves the quality of heuristic functions, large search space still leads to many unnecessary searches. Our second contribution is Regions Discovery Algorithm (RDA), a map clustering technique to partition a grid based map into different categories to reduce search spaces and increase search speed. Our approach reduces the size of search spaces by partitioning a graph into many segments and identifying the segments by their characteristics. By identifying segments in different categories, we can easily eliminate search space, such as rooms, that are not possible (better use needed?) to be part of the optimal solution. Unlike the existing approaches that might result in non-optimal solutions, our experiment results show that RDA guarantees optimal solutions. Our third contribution, the Hierarchical Evolutionary Heuristic A* (HEHA*), further improves the search ability of handling complex pathfinding problems and boosting the search performance, by reducing search spaces and exploiting parallelism techniques. HEHA* combines the strength of EHA* and RDA to reduce search spaces and improve search speed. HEHA* shows that it provides better search performance with less memory consumption. In the pre-processing phase, first HEHA* partitions a graph into different segments and then applies different optimized heuristic functions for each segment to maximize the search performance. During the online process, HEHA* searches on the abstract level first to reduce search area, and exploits parallelism to speed up the search. Fourth, we improve and apply HEHA* to Multi-Agent Pathfinding (MAPF) problems. MAPF is the fundamental problem of many robotic and logistic applications, where the main constraint is that all agents can find the shortest paths while not colliding with each other. While the current trend favors the central controlled system, our approach is to develop a distributed version of HEHA* that can efficiently plan the optimal path for each agent. Such a system requires data sharing and exchanging among the agents, so that each agent can make its own decision without a supervising system. Our experiment results show that the Multi-Agent version of HEHA* maintains a high success rate when the number of agents increases. While EHA* and HEHA* provide a novel approach for heuristic function design, the pre-processing times are not trivial. To boost the performance of the preprocessing steps in EHA* and HEHA*, we propose a FPGA-based reconfigurable hardware accelerator that is not bound to any specific applications as our fifth contribution. Since GA requires many independent processes, it is suitable to implement it in a hardware accelerator to gain maximum performance. We apply the following techniques to enhance performance: deep pipelining, reconfigurable computing, massive parallel processing, and degree of parallelism maximization. Our results show that the FPGA accelerator for EHA* improves the scalability, throughput, and latency