École Nationale Supérieure de
Techniques Avancées

Abstract

Game development has become a long process that requires many professionals working on a project for several months or years. In this context, the reuse of resources is crucial not only to streamline the process, but also to bring consistency to the final product. The problem arises when their behaviors are refined to meet the needs of a specific level without considering that this change may alter their performance on previous levels already tested. We are working on the development of automated testing tools for video games that use artificial intelligence techniques. In particular, we are interested in the following problem: ”given a test specification in a video game environment (e.g., advancing from point A to point B by defeating an enemy and avoiding death), form a policy capable of satisfying this specification as well as systems capable of automatically evaluating this satisfiability at the end of a day of development in order to detect possible variations in the playability of the level”. This is contextualized by the need to maintain exhaustive regression test coverage of the entire game developed to date, considering that as the game grows, it is exponentially expensive to deploy human testers to regularly verify that everything that worked before continues to work as intended. Reinforcement Learning is the AI strategy used for this purpose, often used in numerous problems such as the development of autonomous agents in an optimization context. Although, designing a reward function can be challenging to obtain a desired result. Here, we aimed to use formal temporal logic specification language (TLTL) in order to design an Automaton being then trained to reach an expected goal. To summarize, in this work the goal was to design an entire pipeline for a Video Game designer to specify logical propositions (such as the example given earlier), for creating an autonomous agent able to realize the specifications on Unity3D. A given LTL Formula is first translated in an Automaton which will be used directly in Unity3D to be trained