# Generative Personas That Behave and Experience Like Humans

**Authors:** Matthew Barthet, Ahmed Khalifa, Antonios Liapis, Georgios N. Yannakakis  
**arXiv:** [2209.00459](https://arxiv.org/abs/2209.00459)  
**Date:** 2022-08-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2209.00459v1/

## Key points

- Authors address the challenge of using AI to automatically test games for more immersive and complex game worlds
- Emphasis on generative AI agents, specifically procedural personas, imitating specific playing behaviors through rules, rewards, or human demonstrations
- Proposal to extend behavioral procedural personas to incorporate player experience for a more holistic approach
- Utilization of Go-Explore reinforcement learning paradigm to train human-like procedural personas
- Testing of generative agents on behavior and experience data from over 100 players of a racing game shows distinct play styles and experience responses similar to humans
- Incorporating player experience alongside playing behavior can significantly enhance behavioral exploration in AI-driven game testing scenarios

## Summary

In their paper titled "Generative Personas That Behave and Experience Like Humans," authors Matthew Barthet, Ahmed Khalifa, Antonios Liapis, and Georgios N. Yannakakis address the challenge of using artificial intelligence (AI) to automatically test games for the development of more immersive and complex game worlds. They emphasize the importance of generative AI agents, specifically procedural personas, in imitating specific playing behaviors through rules, rewards, or human demonstrations. However, previous research has primarily focused on behavioral aspects without considering the overall player experience within the game. To bridge this gap, the authors propose extending behavioral procedural personas to incorporate player experience. This enables generative agents to both behave and perceive their gaming environment as humans would. To achieve this goal, they utilize the Go-Explore reinforcement learning paradigm to train human-like procedural personas. The study involves testing these generative agents on behavior and experience data collected from over 100 players of a racing game. The findings indicate that the generated agents exhibit distinct play styles and experience responses similar to those of human players they were designed to emulate. Furthermore, incorporating player experience alongside playing behavior can significantly enhance behavioral exploration in AI-driven game testing scenarios. This innovative approach not only advances AI capabilities in gaming but also holds promise for creating more engaging and realistic virtual environments for players worldwide.