# Learning by Viewing: Generating Test Inputs for Games by Integrating Human Gameplay Traces in Neuroevolution

**Authors:** Patric Feldmeier, Gordon Fraser  
**arXiv:** [2304.06413](https://arxiv.org/abs/2304.06413)  
**Date:** 2023-04-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.06413v1/

## Key points

- Automated test generation is common in many programming domains
- Games present unique challenges due to randomization and hard-to-reach program states
- Neuroevolution combined with search-based testing has shown promise for testing games
- Co-evolutionary search for optimal network topologies and weights can take unreasonably long durations
- Patric Feldmeier and Gordon Fraser propose a novel approach that integrates knowledge about human gameplay behavior into the evolutionary search for networks using traditional gradient descent as a mutation operator
- Experiments on eight diverse Scratch games show that this approach significantly reduces required search time from five hours to only 52 minutes
- This paper presents a promising solution to generating automated tests for games, which could improve game development processes.

## Summary

Automated test generation is a common practice in many programming domains. However, games present unique challenges due to their heavy randomization and hard-to-reach program states. Neuroevolution combined with search-based software testing principles has shown promise for testing games but the co-evolutionary search for optimal network topologies and weights can take unreasonably long durations. Patric Feldmeier and Gordon Fraser have proposed a novel approach that integrates knowledge about human gameplay behavior into the evolutionary search for networks using traditional gradient descent as a mutation operator. Experiments conducted on eight diverse Scratch games demonstrate that this approach significantly reduces the required search time from five hours to only 52 minutes. This paper titled "Learning by Viewing: Generating Test Inputs for Games by Integrating Human Gameplay Traces in Neuroevolution" presents a promising solution to the challenge of generating automated tests for games which could have significant implications in improving game development processes.