Search

Facing the challenge of distributed computing on processing large-scale data, this paper proposes a consensus-based decentralized training method with communication compression. First, the decentralized training method is designed based on the decentralized topology to reduce the communication burden on the busiest agent and avoid any agent revealing its locally stored data. The convergence of the decentralized training algorithm is then analyzed, which demonstrates that the decentralized trained model can reach the minimal empirical risk on the whole dataset, without the sharing of data samples. Furthermore, model compression combined with the error-compensated method is considered to reduce communication costs during the decentralized training process. At last, the simulation study shows that the proposed decentralized training with error-compensated communication compression is applicable for both IID and non-IID datasets, and exhibits much better performance than the local training method. Besides, the proposed algorithm with an appropriate compression rate shows comparable performance with decentralized training and centralized training, while saving a lot of communication costs.

In the Multi-Agent Systems (MAS), graph network topologies play a crucial role in building consensus among the connected agents. Consensus may be achieved on many network graphs using distributed control theory. However, the optimal network topology is not addressed in most of the literature, which is an important part of building stable consensus among networked agents. In this paper, the optimal topology is obtained irrespective of the agent dynamics by using two-dimensional Genetic Algorithm (GA), which is a new approach in this context. Simulation result for agents with first, and second-order linear dynamic is obtained. These results show that the proposed method achieves consensus using the optimal network topology satisfactorily.

In this paper, the consensus problem is investigated for a class of nonlinear multi-agent systems in lower triangular form through dynamic output feedback control. Each agent can only access to the uncertain relative output information from other agents under directed communication. The presence of unknown measurement sensitivities causes the relative output can not be used to construct the observer for each agent directly. For the sake of overcoming this difficulty, a dual-domination gain method is applied to dominate the nonlinear terms and the unknown measurement sensitivities. Then, a new dynamic output feedback consensus prtocol is presented to drive the multi-agent systems to consensus. The upper bound of the unknown measurements sensitivities is also given. At last, we provide two numerical simulation examples to verify the effectiveness of our proposed methods.

In this paper, the leader-following consensus problem is investigated for upper-triangular nonlinear MASs (multi-agent systems) via sampled and delayed output feedback. The sampled outputs transmitted through the communication channels are used to constructed the output feedback controllers directly. Then, the output feedback control signals transmitted through the other communication channels are used to update the MASs. Two kinds of transmission delays, i.e., transmission delays from the MASs to the output feedback controllers and transmission delays from the output feedback controllers to the MASs, make the MASs and the output feedback controllers to be updated at different time instants. A method of interval decomposition is proposed such that the MASs and the output feedback controllers are updated at the same interval. Then, sufficient conditions are presented to ensure the MASs reach leader-following consensus. At last, an example is provided to verify efficiency of the proposed methods.

In this paper, the leaderless consensus problem of multi-agent systems with jointly connected topologies and nonlinear dynamics is considered, in which the nonlinear dynamics are assumed to be non-identical and unknown. The unknown nonlinear dynamics existing in the systems are assumed to be linearly parameterized, and an adaptive design method for leaderless multi-agent systems is presented. By just using the relative position information between each agent and its neighbors, a distributed adaptive consensus control algorithm for the considered systems is proposed, in which the network graphs are jointly connected. Both the global uniform asymptotical stability and the global uniform asymptotical parameter convergence analysis of the adaptive control algorithm are carried out by using adaptive control theory, Lyapunov theory and algebraic graph theory. Finally, an example is given to illustrate the validity of our theoretical results.

In this paper, finite-time adaptive consensus problem is investigated for first-order multiagent systems with unknown nonlinear dynamics. Linearly parameterized method is introduced to model unknown nonlinear dynamics of the systems. By only utilizing the local relative position state information between each agent and its neighbors, decentralized finite-time adaptive consensus algorithms are presented with directed fixed and switching network topologies which satisfy detailed balance condition. Based on classical Lyapunov analysis techniques, both finite-time stability and finite-time parameter convergence are guaranteed by making use of the proposed control algorithms. Finally, the results in Simulations part are presented to validate our main results.

This paper investigates the consensus of second-order multi-agent systems without measuring the velocity states of the agents, where each agent can be either a double integrator or a harmonic oscillator. By utilizing the position information of the agents, a distributed observer-based protocol is proposed to solve the second-order consensus problem of multi-agent systems with or without time delay. The observer-based consensus problem is converted to the stability analysis for a set of second-order quasi-polynomials through coordinate transform and system decomposition. Then, some necessary and sufficient conditions are derived for reaching second-order consensus in multi-agent systems with or without time delay, respectively, and it is shown that the consensus can be achieved if and only if the communication delay is less than a critical value. Simulation examples are given to verify the theoretical analysis.