Loading...
Please wait, while we are loading the content...
Learning One Universal Machine Learning Model for Wi-fi under Diverse Devices and Environments
| Content Provider | Semantic Scholar |
|---|---|
| Author | Liston, Rob Zhu, Xiaoqing Wildfeuer, Herb |
| Copyright Year | 2018 |
| Abstract | Techniques are provided for associating similar devices and environments together so they can be effectively learned. Furthermore, a new device (e.g., smartphone) can be associated quickly with behaviors of other similar observed smartphones to avoid learning from scratch. Since wireless performance depends strongly on device and environments types, any machine learning method also needs to be conditioned on device and environment types. DETAILED DESCRIPTION Applying machine learning to optimize transmission parameters at a Wi-Fi® Access Point (AP) is a promising approach to improving Wi-Fi performance. However, the same transmission parameter may lead to different outcomes for different client types (e.g., between different types of smartphones) and different environments (e.g., conference room versus airport). There are two problems with directly learning separate models at each AP. First, it is ineffective in that contributions from client and environments are tangled and is not transferrable to a new client or a new location for the AP. Second, it is not scalable to train and deploy unique models for every client and environment combination. Instead, methods are needed to characterize client effects and environment effects to allow training one model that can be universally deployed. As such, it is desirable to train a universally deployable machine learning model which accounts for the different influences of devices types (e.g., between different types of smartphones) and environments (e.g., conference room versus airport). Techniques described herein involve capturing the influence of device and environment on wireless network performance via a unified model, which can be trained using the aggregate of data gathered across different sites. These techniques further involve representing device and 2 Tan et al.: LEARNING ONE UNIVERSAL MACHINE LEARNING MODEL FOR WI-FI UNDER DIV Published by Technical Disclosure Commons, 2018 2 5734 environment inputs as embedding vectors, so that proximity in the embedding space generally corresponds to the similarity in device/environment characteristics. This way, the model can easily incorporate new device types via continuous updates. Wireless network telemetry data may be collected from multiple deployment sites and fed to a central location for learning a universal deployment model. Each data entry corresponds to a single packet transmission, and includes the following: observations, actions, performance, client device Identifier (ID) and type, and environment ID. The observations may be of Channel State Information (CSI), Signal-to-Noise Ratio (SNR), etc. The chosen actions may be in the form of transmission parameters such as the number of spatial streams, multi-user versus single-user (MU/SU) mode, etc. The measured performance may include the resulting effective throughput, PHY rate, packet error rate, etc. Client device ID and type may include, e.g., the Media Access Control (MAC) address of the client, prior knowledge pertaining to the device type (e.g., manufacturer name, Operating System (OS) version, etc.), etc. The environment ID may be an identifier of the deployment environment (e.g., MAC address of the AP). One goal may be to train a universally deployable Machine Learning (ML) model for optimizing wireless network performance, e.g., selecting the best transmission parameters to maximize throughput that accounts for the device and environment variabilities. Generally, the ML model can be either a classifier or reinforcement learning policy. It is not scalable to train a ML model with individual device and environment identifiers as direct input, since only a subset of devices is observed at each location. Whenever a new device shows up at a location, the ML model will not be able to make an intelligent decision for this new device-environment combination before it spends some time collecting sufficient data for this new scenario. In contrast, the system described herein can still map the new device to the embedding vector, and its existing knowledge about similar devices at the same location. This way, the system can still perform inference on a new device using an existing model, while quickly updating the model for the new device via continuous training. 3 Defensive Publications Series, Art. 1650 [2018] https://www.tdcommons.org/dpubs_series/1650 3 5734 Figure 1 below illustrates the overall system diagram. Basically, training and inference of the universal model take as input the separately trained device and environment embedding vectors. |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | https://www.tdcommons.org/cgi/viewcontent.cgi?article=2715&context=dpubs_series |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
