Search

In multi-alternative elections, messages in support of or against alternatives spread across online social media. To make decisions, individuals integrate various sources of information, including information from their friends in online social networks. As information may continuously update online, opinions of individuals also become dynamic, making it challenging to predict voting outcomes. In this paper, we propose a novel voting metric named minimal influence gap, which is based on the information structure induced within a social network, to predict voting outcomes in multi-alternative elections. We evaluate this metric as a predictor of voting outcomes across three popular models of voting behaviours: the Maximising Expected Utility model, the Local Dominance model, and the K-Pragmatist model. We test our metric on synthetic networks with a scale-free feature or a community structure, and on a real-world social network. Experimental results demonstrate that the minimal influence gap strongly correlates with voting outcomes in the three models, particularly in the Local Dominance model when voters persist in supporting their favourite alternative.

Multi-agent learning has made significant strides in recent years. Benefiting from deep learning, multi-agent deep reinforcement learning (MADRL) has transcended traditional limitations seen in tabular tasks, arousing tremendous research interest. However, compared to other challenges in MADRL, scalability remains underemphasized, impeding the application of MADRL in complex scenarios. Scalability stands as a foundational attribute of the multi-agent system (MAS), offering a potent approach to understand and improve collective learning among agents. It encompasses the capacity to handle the increasing state-action space which arises not only from a large number of agents but also from other factors related to agents and environment. In contrast to prior surveys, this work provides a comprehensive exposition of scalability concerns in MADRL. We first introduce foundational knowledge about deep reinforcement learning and MADRL to underscore the distinctiveness of scalability issues in this domain. Subsequently, we delve into the problems posed by scalability, examining agent complexity, environment complexity, and robustness against perturbation. We elaborate on the methods that demonstrate the evolution of scalable algorithms. To conclude this survey, we discuss challenges, identify trends, and outline possible directions for future work on scalability issues. It is our aspiration that this survey enhances the understanding of researchers in this field, providing a valuable resource for in-depth exploration.

Abstract Coordinating multiple agents to optimize an objective has several real-world applications. In areas such as disaster rescue, environment monitoring and the like, mobile agents may be deployed to work as a team to achieve a joint goal. Recently, multi-agent problems involving mobile sensor teams have been formalized in the literature as DCOP_MSTs. Under this class of problems, DCOP algorithms are applied to enable agents to coordinate the assignment of their physical locations as they jointly optimize the team objective. In DCOP_MSTs, the environment is dynamic, and agents may leave or join the environment at random times. As a result, a predefined interaction topology or graph may not be useful over the problem horizon. Therefore, there is a need to study methods that could facilitate agent-to-agent interaction in such open and dynamic environments. Existing methods require reconstructing the entire graph upon detecting changes in the environment or assume a predefined interaction graph. In this study, we propose a dynamic multi-agent hierarchy construction algorithm that can be used by DCOP_MST algorithms that require a pseudo-tree for execution. We evaluate our proposed method in a simulated target detection case study to show the effectiveness of the proposed approach in large agent teams.

In multi-agent systems, the Dynamic Distributed Constraint Optimisation Problem (D-DCOP) framework is pivotal, allowing for the decomposition of global objectives into agent constraints. Proactive agent behaviour is crucial in such systems, enabling agents to anticipate future changes and adapt accordingly. Existing approaches, like Proactive Dynamic DCOP (PD-DCOP) algorithms, often necessitate a predefined environment model. We address the problem of enabling proactive agent behaviour in D-DCOPs where the dynamics model of the environment is unknown. Specifically, we propose an approach where agents learn local autoregressive models from observations, predicting future states to inform decision-making. To achieve this, we present a temporal experience-sharing message-passing algorithm that leverages dynamic agent connections and a distance metric to collate training data. Our approach outperformed baseline methods in a search-and-extinguish task using the RoboCup Rescue Simulator, achieving better total building damage. The experimental results align with prior work on the significance of decision-switching costs and demonstrate improved performance when the switching cost is combined with a learned model.