Extending Goal-Oriented Specification for Deep Reinforcement Learning to High-Dimensional State Spaces
This work extends goal-oriented specification for deep reinforcement learning to high-dimensional state spaces by introducing an intuitive feature space for defining goals independently of raw state representations. It also adds a dedicated training space, that is constructed from one or both representations, improving training in complex environments. Case studies in Pendulum and Pong show that the approach enables effective goal-oriented specification beyond low-dimensional settings.