Multi-objective reinforcment learning in Brax

Abstract

This thesis looks at how PG-Multi objective reinforcement learning can be used to solve control when there is more than one goal. PG-MORL was tested on HalfCheetah and Hopper environments where agents had to objectives to achieve move faster while using less energy . These two goals often work against each other, so the challenge was to find good trade-offs instead of just optimizing one. The idea was to train multiple policies using fixed weight combinations for speed and energy. This way, each policy would reflect a different balance. In the end, the algorithm was able to build up an archive of diverse behaviors. Some policies moved quickly, some were energy-efficient, and some found a solid middle ground. The results showed that PG-MORL works in different environments without needing complex adjustments. It’s simple to set up and gives you a useful set of options for when your goals aren’t fixed. Even with a basic scalarization method, the approach was able to produce meaningful results.