Overview
Explore how deep reinforcement and imitation learning can revolutionize playtesting and NPC creation in game development. Delve into Unity's Jeffrey Shih's 2020 GDC Virtual Talk, which covers the most common use cases, effective approaches like domain randomization and guided demonstrations, and strategies to mitigate costs. Gain insights into real-world applications through case studies such as Carry Castle and Source of Madness, learning valuable lessons on visualizations, reward balancing, and action handling. Discover the potential of machine learning to scale and enhance game testing and character development processes.
Syllabus
Intro
Why Unity @ GDC ML Summit
How are studios using DRL?
Most common use case
Reinforcement learning in a nutshell
Test new levels or content using RL
A few effective approaches
Domain randomization
Using demonstrations to guide RL
How do we mitigate cost?
Increasing sample throughput
Increasing sample efficiency
Using RL for testing - final thoughts
Carry Castle
Challenges
RL setup for Source of Madness
Structuring the proper rewards
Lessons Learned - Visualizations
Lessons Learned - Balancing Rewards
Lessons Learned - Handling of Actions
Taught by
GDC