Local hyperparameter optimization for causal local states

Develop efficient strategies to optimize the separate hyperparameter configurations of the multiple local models in the causal local states framework without exhaustive search, rather than relying on a shared configuration across all neighborhoods.

Background

The causal local states framework trains multiple local predictive models, one for each core node and its inferred neighborhood. These local models may require different hyperparameters because the neighborhoods can represent heterogeneous dynamical regimes. In the reported experiments, a single shared hyperparameter configuration is used for all neighborhoods, but the authors note that this is unlikely to be optimal. The unresolved problem is therefore to devise computationally efficient, non-exhaustive procedures for tuning hyperparameters locally while retaining the scalability of the framework.