Simplifying Visual Data to Improve Reinforcement Learning with a Classifier System

What is it about?

Learning Classifier Systems (LCSs) are a form of rule-based artificial intelligence that can be understood by both humans and mathematically. A well-known variant, XCSF, can solve many learning tasks but struggles with visual input such as images because the data is too complex. In our research, we combined XCSF with variational autoencoders (a type of neural network that compresses images) to make reinforcement learning from visual data possible. We also asked whether simpler methods could work, and found that even basic image downscaling allowed XCSF to learn visual tasks. This shows that classifier systems seem to be more powerful than expected, and opens the door to applying them to a wider range of problems.

Featured Image

Why is it important?

Many popular reinforcement learning benchmarks, as well as real-world AI tasks, involve images or other large, complex inputs. Our research shows how a Classifier System can learn from these larger inputs without modifying the system itself. This could broaden the range of problems this type of rule-based AI—understandable by both humans and mathematically—can tackle, making it potentially more practical for both benchmark problems and real-world applications.