A Control Strategy for Multi-robot Systems Based on a Gradient-Based Differential kWTA Network

The k-winners-take-all (kWTA) mechanism is widely used to address multi-agent decision-making and coordination problems. Existing kWTA networks in multi-agent tracking often suffer from slow convergence and delay-induced errors. To address these issues, this paper proposes an improved gradient-based differential k-winners-take-all network, transforms the kWTA problem into a constrained dynamic quadratic programming problem and solves it using a neural network to accelerate convergence and eliminate lagging errors. Specifically, the network incorporates time-derivative information to eliminate lagging errors caused by dynamic inputs and designs a special Maxout activation function to further accelerate the network’s convergence speed. Theoretical analyses and numerical experiments demonstrate that this method can effectively improve convergence performance and eliminate lagging errors. In addition, multi-agent tracking simulations demonstrate that integrating the GD-kWTA network with a consensus estimator effectively resolves challenges such as agent competition and communication constraints, validating the method’s effectiveness and practicality in distributed systems.