A common challenge faced by all decentralized multiagent systems (MAS) is the exploration-exploitation dilemma. This stems from the fact that gathering new information about the environment (ie, exploration) and making use of currently known information (ie, exploitation) tend to be mutually exclusive activities. While heavily biasing an MAS towards exploration would allow large amounts of data to be gathered, it would be prevented from fully benefiting from this information. Conversely, over-exploitation may yield fast convergence times but can also result in agents being trapped in local optima or being unable to adapt to dynamic environments. Therefore, maximizing the performance of a system, especially when operating in dynamic environments, requires some sort of regulation of the amount of exploration and exploitation carried out by an MAS. In this work, we outline how the strategy developed in our work (Kwa et al., 2021) allows for the regulation and control of a swarm’s exploration and exploitation dynamics (EED) within the context of tracking fast-moving evasive and non-evasive targets.

Previously, Esterle and Lewis (2020) have shown that increasing the level of inter-agent communications improved the overall tracking performance of the system while tracking slow-moving targets. However, it has been demonstrated that this is not the case when tracking fast-moving nonevasive targets (Kwa et al., 2020). This is because higher levels of inter-agent connectivity leads to over-exploitation of a target’s positional information, causing agents to lose track of the target once they have been outrun. Similarly, low levels of connectivity result in too much exploration and insufficient exploitation, preventing agents from effectively tracking the target. As such, an optimal level of connectivity exists at which the level of exploration and exploitation carried out by the MAS was relatively balanced, maximizing its tracking performance. Similar optimal levels of connectivity maximizing system performance were found in other scenarios such as in obtaining a dynamic consensus (Mateo et al., 2019) and in a collaborative stick pulling task (Hamann, 2018).