Modular Learning for Multiple Wireless Policies with PE Properties

Recognizing that designing a specific model for learning each wireless policy (i.e., a mapping from known environment parameters to decision variables) leads to an unsustainable proliferation of specialized models, recent trends have shifted towards adapting a large deep neural network (DNN) for multiple policies. However, these large models lack structural domain knowledge and are hard to pre-train and fine-tune at the wireless edge. To facilitate efficient learning of multiple policies with a single DNN, we conceive a modular approach in this work by leveraging structured knowledge of wireless policies - specifically, permutation equivariant (PE) properties that widely exist in wireless policies. We design three PE-aware modules, two of which are in a Transformer-style. These modules serve as reusable building blocks for assembling compact DNNs applicable to wireless policies with different PE properties. To guide how to compose the modules appropriately for boosting the learning efficiency, we analyze the PE-constrained hypothesis space. Inspired by the reusability of the modules, we propose PE-MoFormer, a compositional DNN capable of learning multiple wireless policies, including channel estimation, signal detection, precoding, coordinated beamforming, and power allocation, with a small model size and strong generalizability to problem scales and channel conditions. Simulations demonstrate that the proposed PE-MoFormer outperforms non-PE-aware large models in both learning efficiency and inference time, offering a promising direction in structured learning for multiple wireless policies with PE properties.