A generative simulation platform for multi-agent systems with incentives

Multi-agent systems have attracted much attention in the recent years due to their capabilities to handle complex and computation-heavy tasks and compatibility with incentive schemes. Considering the difficulty of creating an actual prototype and environment for evaluation, a simulation platform is a cheap and efficient way in analyzing and testing, prior to real environmental implementations. Existing simulators for multi-agent systems are inadequate to analyze the effects of different customized incentive schemes on agents' behavior patterns due to two reasons: 1) They lack the functionality to support various types of complex incentives, e.g., mixture of monetary incentives and non-monetary incentives, which influences agents' behaviors explicitly and implicitly; 2) They are not able to emulate heterogeneous agents' realtime behaviors that are influenced by complex incentives and deviate from their original behavior patterns shown in historical traces. In this paper, we focus on mobile agents that can move in a patio-temporal space, and we present a physical knowledge aided multi-agent simulation platform considering the influence of both direct and indirect incentives unified through a general utility-driven agent reaction function. The behaviors of agents are then emulated in three behavioral models: myopic, semi-myopic, and farsighted, by varying the assumption of agents in maximizing their utilities and integrating the physical knowledge and historical mobility patterns. We finally examine the effectiveness of the platform in incentivizing vehicle agents to optimize the final distribution of the agents through a ride-sharing vehicle experimental scenario. The emulated agents' behaviors can also be collected into data traces for analyzing other patterns of the agents.