| Abstract |
In multi-domain operations, different domains get different modalities of input signals, and as a result end up training different models for the same decision-making task. The input modalities could be overlapping with each other and could be of different natures (continuous versus discrete). This leads to situations in which specific machine learning models may be more effective in learning from data of a certain type than others (e.g., differentiable learning versus symbolic learning), and models created in one domain may be reusable partially for the decision-making task being conducted in other domains. In order to share the knowledge embedded in different models trained independently in each individual domain, we propose the concept of hybrid policy-based ensembles, in which heterogeneous models from different domains are combined into an ensemble whose decision-making operation is controlled by policies specifying how these models can be combined to perform the operation. We show how these policies can be expressed based on several factors characterizing the training datasets, e.g., feature distributions and feature importance. In the paper we also report performance results of these hybrid policy-based ensembles on an intrusion detection dataset and a human activity recognition dataset |