Research On Infrastructure, Organizations And AI Algorithms For Multi-Agent Simulation

In recent years, simulation communities have begun to focus large scale intelligent, concurrent complex systems with multiple organization structures. Qualitative and quantitative research on these systems needs to integrate domains of computer simulation, system theory and artificial intelligence (AI). Multi-agent based simulation (MABS) builds target system's high level model with system theory and multi-agent system modeling methodology, then implements and run its computational model adopting simulation software and hardware infrastructures facilitating agent based modeling. It can handle the nonlinear, interactive and emergent features of complex system, and has turn out to be the uppermost and effective mean for the research of complex system.Now MABS is moving from the explorative stage to the applicable one, and there are many theoretic and practical problems need to be studied, such as lack of generic simulation infrastructure with reasonable scalability, inadequacy in deep-seated questions like organization modeling for simulation, and application limitation in small number of traditional domains. To address these problems, this paper firstly carries out study of constructing and perfecting the theory of MABS and general infrastructure, then concentrates on providing simulation-oriented computational models for analyzing temporal/spatial structures of agent organizations, and extends the multi-agent simulation to further application domain such as military simulation. The contributions of this paper are listed in detail as follows:1. Multi-agent based simulation kernel model with semantic duration and organization information extension. It is essential to set up formal model and design strategies according to patterns of agent's behavior and interaction to build up simulation. After reviewing traditional researches, we propose a kernel model which lay a formal basis for designing MABS's infrastructure and strategy. Considering temporal constraints on agent's executive model under simulation specification, we propose a semantic duration model which mapping time effects of the behavior semantic to appropriate simulation durations. According to specifications for modeling agent society structures and organizations, we propose two extended organization models for simulation which are respectively based on the implications of structure and constraint of the organization concept.2. Agent-based framework for complex system simulation and its implementation. We propose a flexible agent based distributed framework for complex system simulation—FFCAS. FFCAS uses complex system modeling framework containing external events and macro rules specification. Functional components of FFCAS are deployed according to distributed interactive simulation architecture and "Model-View-Controller"design pattern. FFCAS is implemented as a generic distributed complex system simulation platform—Advanced JCass. Inter-domain applications, namely the grassland ecological system, public attitude system and public scientific literacy system, have been implemented for demonstration. Empirical results show that FFCAS has enough flexibility, reusability and scalability in handling complex system simulation.3. Simulation-oriented agent organization model and its implementation. Aiming at describe the spatial structure evolvement of complex system, studies have been focused on two typical organization models: one is social network and the other is role-based paradigm. Social network is applied to solve the knowledge optimization problem in public scientific literacy system. An approximate algorithm using heuristic based on the maximum propagating influence of vertices and its provable guarantee of (1-1/e) is presented. Then a role-based modeling and instantiating framework for simulation organization is proposed, and is used to specifying agent organization structure and dynamics in cyber warfare simulation as a demonstration.4. Simulation-oriented AI algorithms for multi-agent. As to bridges researches of simulation and AI, learning and collaboration are used as analysis cases for they are common focus of simulation and distributed AI. Firstly existing AI techniques in simulation are surveyed. An opponent learning algorithm for certain MABS scenario is proposed, and has the feature that rewards of agent's behavior under multi-player adversarial environment are determined not only by its own choice but also its opponent's preferences. To relax the NP-hard problem of planning in multi-agent simulation, a simulation-oriented hierarchical multi-agent problem solving algorithm with instance demonstration is presented, which is based the fact that concurrent problem solving among multiple agents using hierarchical abstraction can reduce search to from exponential to logarithmic complexity.5. Multi-agent based simulation on networked air defense warfare. Air defense warfare simulation is an extended multi-agent simulation application beyond traditional domains such as social science issues. To match the intention of study situation sharing and forces coordination in networked air defense warfare, multi-agent modeling, agent organization and intelligent algorithms are applied to build up the simulation networked air defense warfare. Also an algorithm is presented to handle the weapon-target assign problem in air defense warfare. Finally, an open MABS framework——integrated simulation environment for networked air defense coordinate combat, is proposed according to domain requirements.