Hierarchical model-based policy optimization: from actions to action sequences and back
This work addresses hierarchical policy optimization for reinforcement learning, but it appears incremental as it builds on existing model-based methods and is tested only on toy problems.
The paper tackled the problem of hierarchical model-based policy optimization by developing a normative framework using second-order methods in state-action paths, resulting in a natural path gradient that updates policies based on long-range correlations. In simulation, they demonstrated that this approach prioritizes local policy updates reflecting intuitive state-space hierarchies in toy problems.
We develop a normative framework for hierarchical model-based policy optimization based on applying second-order methods in the space of all possible state-action paths. The resulting natural path gradient performs policy updates in a manner which is sensitive to the long-range correlational structure of the induced stationary state-action densities. We demonstrate that the natural path gradient can be computed exactly given an environment dynamics model and depends on expressions akin to higher-order successor representations. In simulation, we show that the priorization of local policy updates in the resulting policy flow indeed reflects the intuitive state-space hierarchy in several toy problems.