An Efficient Sampling-Based Algorithms Using Active Learning and Manifold Learning for Multiple Unmanned Aerial Vehicle Task Allocation under Uncertainty

Abstract

This paper presents a sampling-based approximation for multiple unmanned aerial vehicle (UAV) task allocation under uncertainty. Our goal is to reduce the amount of calculations and improve the accuracy of the algorithm. For this purpose, Gaussian process regression models are constructed from an uncertainty parameter and task reward sample set, and this training set is iteratively refined by active learning and manifold learning. Firstly, a manifold learning method is used to screen samples, and a sparse graph is constructed to represent the distribution of all samples through a small number of samples. Then, multi-points sampling is introduced into the active learning method to obtain the training set from the sparse graph quickly and efficiently. This proposed hybrid sampling strategy could select a limited number of representative samples to construct the training set. Simulation analyses demonstrate that our sampling-based algorithm can effectively get a high-precision evaluation model of the impact of uncertain parameters on task reward.

Combining an Extended SMAA-2 Method with Integer Linear Programming for Task Assignment of Multi-UCAV under Multiple Uncertainties

“…l ← 1 ∆ρ (l) ← ∞ while ∆ρ (l) ≥ ζ do compute λ by (33) l ← l + 1 for k ← 1 to K do compute Dev S (k) , S (0) by (22), (23), (31), (32) and (34) end for update ρ (l) by (35) update ∆ρ (l) by (36) end while return S (0)…”

“…The authors then solved the task assignment problem in the framework of a two-stage stochastic programming model using a genetic algorithm [3]. A similar way of dealing with uncertain parameters can be found in Reference [22], where the task duration time was considered to be subject to a uniform distribution.…”

Decentralized task allocation for multiple UAVs with task execution uncertainties

“…In specific multi-UAV task allocation problems, the uncertain duration often have their own distribution, which can be obtained from historical data, surveys or theoretical analysis [10]. While the true value of is unknown, it is assumed that a likelihood model of the uncertainty parameter is known beforehand, obeying a statistical probability distribution:…”

“…Lemma1: For reward function( 9), if is the optimal order of tasks in the task set , for any task in , it would satisfies (10) where represents the task duration of task .…”

“…In an effort to retain the outcome with underlying uncertainties, Ponda [9] then presents the robust extension to CBBA using the integral of expected reward instead. However, analytically computing these robust scores remains cumbersome [9] [10].…”