Location determination in wireless networks using machine learning

A network controller comprising:

• a processing resource;
• and
• a non-transitory machine-readable storage medium storing instructions executable by the processing resource, wherein the processing resource is configured to execute one or more of the instructions to:
• receive, from an access point (AP), a runtime signal strength metric indicative of wireless signal power between the AP and a client device in a network infrastructure;
• receive, from the AP, a runtime environmental condition metrics corresponding to one or both of the AP and the client device;
• determine a coarse distance between the AP and the client device based on the runtime signal strength metric;
• and
• determine a refined distance between the AP and the client device based on the coarse distance and the runtime environmental condition metrics using a machine learning (ML) model, wherein the ML model is trained to infer the refined distance based at least on a set of test environmental condition metrics corresponding to a set of network devices installed at known locations in the network infrastructure and one or more test signal strength metrics reported by the set of network devices.

A computer-implemented method comprising:

• receiving a test signal strength metric between a pair of network devices positioned at known locations in a network infrastructure;
• receiving an actual distance between the pair of network devices;
• determining a preliminary distance between the pair of network devices based on the test signal strength metric;
• finetuning the preliminary distance to generate a finetuned preliminary distance using a neural network, a set of test environmental condition metrics, and a set of neural network weights, wherein the set of test environmental condition metrics comprises measures of one or more of humidity, temperature, pressure, or signal absorption by physical objects;
• updating the set of neural network feature weights to generate a set of updated neural network weights based on a difference between the actual distance and the finetuned preliminary distance;
• and
• recording the set of updated neural network weights.