Machine Learning-Based Open Framework for Multiresolution Multiagent Simulation

摘要

M&S of systems, their dynamic structures and particularly the behaviour of internal component objects, should be performed at the level of detail which is adequate to the problem and modelling purpose defined. In the scenarios related to the complex world strictly one level is insufficient - there is necessary to build a multi-resolution model that represents structures and actions at different levels of detail. This is the main and direct reason for the application of the Multi-resolution Agent Model (MrAM) approach in the simulation with functions for a state's transformation (aggregation/disaggregation). It is common practice to implement methods of resolution adaptation in such a way that they are completely closed in the compiled program code. Meanwhile, the multiplicity of different possible scenarios regarding group and individual behaviours indicates that there are necessary software constructions enabling the end user to create both new, open models of behaviours and algorithms for aggregation/disaggregation of the state. Moreover, the environment surrounding agents influences target states differently at different moments in time. The article proposes an approach to determining the consensus state of agents with the use of machine learning methods. The consensus between agents according to the appropriate approach, depending on the conditions and state, will be a generalized method adaptable to the environment of agents. We propose the reinforcement learning model as a multiagent game in order to achieve HRE state and thus complete disaggregation. To meet the requirements, the original Java-based framework for hybrid simulation (discrete, event-based and continuous) with the ability to model the object as an agent at multiple levels, automatic triggering of updates on all modelled levels, and Groovy-based scripts. The scripting technology is integrated with the standalone Java software and enables the implementation of behaviors and state transformations that are really open in scripts. The article presents the proposed framework solution on the example of an autonomous system model composed of many cooperating objects. We share our experiences related to the extension of the "SymSG Border Tactics" simulation environment dedicated to CAX exercises in The Poland Border Guard.