What are the most important open challenges?
- Open-ended action spaces
- Large action spaces
- Large numbers of agents
- Multi-agent partially observable games
- Synthesizing complex movements for humanoid characters
- Learning general behavior
- Hierarchical reasoning
- Strategy learning for general games
-
Explainable AI
- Optimal RL: Generalization, transfer learning, sample efficiency, multi-task, model-based, etc.
- Sample efficiency of RL methods, high variance and robustness to underlying environment changes
-
Learning with a few samples and transfer learning
- Learning world models
-
Learning forward model abstractions
- Content evaluation
- Evaluating computational creativity applications
- How to get software to write software in creative ways that supplement human coders - this will revolutionise game design (and a lot of other areas)
- Creating AI-assisted designers that actually collaborate with humans
-
Procedural content generation for commercial games
- How the use of machine learning models to customise game control schemes can affect the player experience
