Generative adversarial networks-based dynamic multi-objective task allocation algorithm for crowdsensing

Task allocation of large-scale and widely-distributed mobile users in a crowdsensing system is a challenging issue, especially on time-varying sensing requirements and available users. In order to shrink the search space, the sensing area is evenly segmented into several subarea, and the tasks in a subarea are allocated to the users preferring it. Following that, a dynamic multi-objective task allocation model is formulated, and an end-to-end algorithm based on generative adversarial network (GAN) is proposed as its problem-solver, with the purpose of quickly generating the Pareto-optimal allocation schemes. Specifically, after decomposing multi-objective problem into a set of scalar subproblems, each subproblem in each subarea is modeled as a neural network, in which the outputs represent the probabilities of users assigned to tasks. In order to produce better solutions and jump out local optima, a new loss function is designed for the discriminator. Besides, a rapid response mechanism is developed to timely adjust the structure of generative nets once a dynamic event occurs, avoiding the time-consuming re-training. Extensive experiments have been conducted to analyze the performance of proposed algorithm. The statistical results expose the competitiveness of proposed algorithm in terms of running time and performance.