Learning End-User Behavior for Optimized Bidding in HetNets: Impact on User/Network Association

TL;DR

This paper investigates how end-user behavior modeled by Prospect Theory affects bidding strategies and user association in HetNets, proposing a learning-based framework to optimize SP bidding under uncertainty.

Contribution

It introduces a novel two-stage learning framework combining SVM and reinforcement learning to improve SP bidding strategies considering user decision deviations.

Findings

Underweighting service guarantees increases bid rejection rates.

Proposed framework improves SP utility and service rates.

Simulation validates model efficiency and effectiveness.

Abstract

We study the impact of end-user behavior on service provider (SP) bidding and user/network association in a HetNet with multiple SPs while considering the uncertainty in the service guarantees offered by the SPs. Using Prospect Theory (PT) to model end-user decision making that deviates from expected utility theory (EUT), we formulate user association with SPs as a multiple leader Stackelberg game where each SP offers a bid to each user that includes a data rate with a certain probabilistic service guarantee and at a given price, while the user chooses the best offer among multiple such bids. We show that when users underweight the advertised service guarantees of the SPs (a behavior observed under uncertainty), the rejection rate of the bids increases dramatically which in turn decreases the SPs utilities and service rates. To overcome this, we propose a two-stage learning-based optimized bidding framework for SPs. In the first stage, we use a support vector machine (SVM) learning algorithm to predict users' binary decisions (accept/reject bids), and then in the second stage we cast the utility-optimized bidding problem as a Markov Decision Problem (MDP) and propose a reinforcement learning-based dynamic programming algorithm to efficiently solve it. Simulation results and computational complexity analysis validate the efficiency of our proposed model.