Multi-Agent Systems for Collaborative Inference Based on Deep Policy Q-Inference Network

Abstract

This study tackles the problem of increasing efficiency and scalability in deep neural network (DNN) systems by employing collaborative inference, an approach that is gaining popularity because to its ability to maximize computational resources. It involves splitting a pre-trained DNN model into two parts and running them separately on user equipment (UE) and edge servers. This approach is advantageous because it results in faster and more energy-efficient inference, as computation can be offloaded to edge servers rather than relying solely on UEs. However, a significant challenge of collaborative belief is the dynamic coupling of DNN layers, which makes it difficult to separate and run the layers independently. To address this challenge, we proposed a novel approach to optimize collaborative inference in a multi-agent scenario where a single-edge server coordinates the assumption of multiple UEs. Our proposed method suggests using an autoencoder-based technique to reduce the size of intermediary features and constructing tasks using the deep policy inference Q-inference network’s overhead (DPIQN). To optimize the collaborative inference, employ the Deep Recurrent Policy Inference Q-Network (DRPIQN) technique, which allows for a hybrid action space. The results of the tests demonstrate that this approach can significantly reduce inference latency by up to 56% and energy usage by up to 72% on various networks. Overall, this proposed approach provides an efficient and effective method for implementing collaborative inference in multi-agent scenarios, which could have significant implications for developing DNN systems.