Combining admission and modulation decisions for wireless embedded systems

Wireless communication is increasingly being used to federate embedded devices in a variety of distributed systems application domains, ranging from wireless sensor networks to the emerging "Internet of Things (IoT)." Since such embedded devices are tightly coupled both with their environments and with each other through their wireless communication channels, both variations in their environments and the system's need to respond (sometimes rapidly) to those variations may produce (1) the need for such devices to communicate and (2) with it the potential for channel contention to arise, dynamically at run-time. Thus, how wireless channels among the embedded devices are allocated and managed in these systems may significantly influence both communication-specific quality-of-service (QoS) properties (such as message throughput) and broader QoS properties (such as timeliness of system responsiveness) that depend on them. A growing body of research has focused on managing different aspects of wireless communication, but has done so mainly in an ad hoc manner, with respect to individual aspects rather than multiple aspects and their potential interactions. Even less attention has been paid to formal methods for assessing how combinations of aspects may influence communication performance, and how to characterize, adapt to, and exploit their combined effects, which is essential to address the challenges noted above. To overcome these limitations of the current state of the art, this paper makes three main contributions to wireless communication for distributed embedded systems with QoS constraints. First, it shows how a basic but fundamental set of channel admission and modulation decisions can be combined within a single Markov decision process (MDP) model to optimize (in expectation) objectives such as message throughput, even with stochastic arrival and interference characteristics. Second, it identifies regular structure in the value-optimal policies generated off-line from these models, which forms the basis for efficient and accurate heuristics suitable for on-line use. Third, it shows how single-and multi-variable regression techniques can be used to characterize key parameters that govern such regular structure, which then are used to instantiate those heuristics.