Accelerate Personalized IoT Service Provision by Cloud-Aided Edge Reinforcement Learning: A Case Study on Smart Lighting
To enhance the intelligence of IoT devices, offloading sufficient learning and inferencing down to the edge environment is promising. However, there are two main challenges for applying the cloud generated model in the edge environment. On the one hand, t
- PDF / 1,423,763 Bytes
- 16 Pages / 439.37 x 666.142 pts Page_size
- 65 Downloads / 143 Views
Abstract. To enhance the intelligence of IoT devices, offloading sufficient learning and inferencing down to the edge environment is promising. However, there are two main challenges for applying the cloud generated model in the edge environment. On the one hand, the input may vary on dimensions or cover different situations that the cloud hasn’t met. On the other hand, the model’s output might not satisfy the given user’s personalized preference. To make full use of the cloud generated model in the edge environment for accelerating personalized service provision, we propose cloud-aided edge learning. Unlike current federated learning and transfer learning, we focus on knowledge fusion in edge decision making and try to build the supplement/correction model. We take the personalized service provision in a smart lighting system as an example, design and implement the related deep reinforcement learning model, and take experiments based on the data generated on the open software DAILux to show our approach’s effectiveness and performance. Keywords: Edge intelligence · Edge-cloud collaborated learning · Personalized service provision · Smart lighting · Deep Reinforcement Learning (DRL)
1
Introduction
The Internet of Things (IoT) [3,20] enables all kinds of real-world objects (including human beings) to be connected to the cyber world. Considering the characteristics of human-in-the-loop, providing personalized IoT services efficiently and transparently turns to be essential. Recently, applying machine learning to speed up personalization becomes a promising way[22,36], which can extract useful knowledge from interactions happening in the physical world to produce proper reactions. To process the continuously generated IoT data efficiently, it needs a powerful data center with enough storage and computing resources. Although cloud computing is an excellent platform to handle the enormous IoT data, pushing all the raw data to the cloud is inefficient in response latency, network c Springer Nature Switzerland AG 2020 E. Kafeza et al. (Eds.): ICSOC 2020, LNCS 12571, pp. 69–84, 2020. https://doi.org/10.1007/978-3-030-65310-1_6
70
J. Na et al.
bandwidth cost, and possible privacy concerns [8,23]. To solve these problems, edge computing [28], also known as fog computing [4], is becoming the right solution and get more attention in both research and industry domain. By offloading sufficient training and inferencing down to the edge environment, edge intelligence would be enhanced to satisfy users’ personalized needs more efficiently while protecting privacy [19,27,32,37]. Combining both cloud computing and edge computing advantages to offer flexible edge-cloud collaboration gets more attention [5,27,36]. Existing studies usually focus on the underlying mechanisms of edge-cloud collaboration. However, there are more challenges to accelerate personalized service provision through deep learning. For example, data achieved by the edge node might be different from the generic dataset used to generate the global model. It does not only refer t
Data Loading...
