A Deep Reinforcement Learning Approach for Pathfinding in Computer Games

First International Symposium in Graduate Researches on Data Science, 02 Aralık 2022, ss.34

One of the biggest challenges of game development is to produce a pathfinding algorithm that both

produces satisfactory realistic movement results and can solve different scenarios in worlds created by game

developers' unlimited imagination. Furthermore, since the games are programs for the end user, it is desired that

the systems they contain use as little computer resources as possible and be developed as quickly as possible in

terms of cost. With the advances in software, lightweight methods have been developed that can provide general

map coverage in this regard. Although existing solutions can produce strong answers to the problem, they also

contain some chronic problems that take a long development time to solve. Existing solutions work very well on

maps that are continuous and can be navigated by moving from start to finish. However, it cannot find a solution

in cases where various obstacles must be overcome by various movement mechanics such as jumping, flying or

dash are used other than walking. Developers must manually assign links to meshes, significantly prolonging the

game development process on the maps with such features. At the same time, since the connected links are

established manually, the movement of the object moving on the link does not seem natural. The focus of this

study is to create a system that will generate a node network by using artificial neural networks and deep

reinforcement learning to overcome the difficulties of existing pathfinding algorithms. Finally, a system that is

fast and uses less resources is aimed for the end user, since artificial neural networks will not be used during the

build phase. Some promising results have been obtained so far, which shows that the alternative methodology

proposed in this study could be a useful alternative for game developers.