Force Protection Call 4 A-0938-RT-GC EUSAS European Urban Simulation for Asymmetric Scenarios Scalarm: Massively Self-Scalable Platform for Data Farming © EADS 2010 – All rights reserved Agenda 1. 2. 3. 4. 5. 6. 7. Introduction to Data Farming in the EUSAS project The problem of Data Farming scale Overview of Scalarm Architecture of Scalarm Resource management Scalarm applications Conclusions © EADS 2010 – All rights reserved 2 Goals of data farming in EUSAS However, As Result During Each Agents Moreover, a result, agent of training can the ifgroups we many also can mission look sessions be be scenarios may parameterized! divided closer, cannot have soldiers the into of leaders. bemission simulation groups! predicted perform areaprocess simply possible mission on is(scenario). but the more soldiers basis complicated of are the able fewthan serious to it seems. execute game executions mission only of the a few scenario. times during the training session. Data farming allows analysis of missions that have several different parameter combinations (possibly billions). 1. group that loots: Leader of group 1: •radius of influence - 10, •prestige - 30, •… •group size - 30, •aggression of members Agent 1: 10, •readiness for for aggression •readiness ofaggression members -- 510, •anger - 3, 2. group •fear - 12, that is prone to •….violence •group2:size - 12, Agent •aggressionforof members •readiness 10, aggression - 10, •readiness •anger - 10,for aggression of members - 25 •fear - 2, •…. In EUSAS project data farming provides: • Identification of dependences between input parameters and simulation result (described by measures of effectiveness). • Comparison of behaviour models/soldiers strategies. • Selection of input parameters for training sessions © EADS 2010 – All rights reserved 3 Introduction to Data Farming in the EUSAS project • Data Farming, in general, enables discovery of useful insights in studied phenomena by providing large amounts of data for analysis. • In the context of EUSAS, Data Farming is utilized to study agents’ behaviour in various scenarios in order to verify different engagement strategies. • EUSAS developed a novel system, called Scalarm, to facilitate conducting large Data Farming experiments with heterogeneous computational infrastructure. © EADS 2010 – All rights reserved 4 The problem of Data Farming scale – simple scenario • Two groups: 1. Looters group 1 2. Group that is prone to violence • Two informal leaders: • • 2 Many input parameters: • • soldiers do not know who they are group sizes, leader prestige, readiness for aggression… Many monitored MoEs: • • escalation, anger, number of killed or injured agents… Presented application of data farming system: • Identification of dependencies between input parameters and simulation results (described by Measures of Effectiveness). © EADS 2010 – All rights reserved 5 The problem of Data Farming scale • Simple simulated scenario can include: • 2 individual agents representing group leaders (each with 22 parameters) • 2 groups of agents (each group with 24 parameters) • • • • => 92 different parameters for a single scenario Let’s suppose we want to check only 2 values for each parameter => 2^92 different simulations Let’s suppose a single simulation runs only for 1 second on average => 157,019,284,536,451,074,949 compute years We need to filter input parameter combinations even more and have a lot of computing power at the backend. © EADS 2010 – All rights reserved 6 Scalarm goals • • • • • Simulating complex phenomena with multiple input parameters by running various types of simulation applications, e.g. multi-agent, optimization, etc. Self-scalable platform adapting to particular problem size and different simulation types Exploratory approach for conducting Data Farming experiments Supporting online analysis of experiment partial results Running on Cloud, Grid and private cluster infrastructures © EADS 2010 – All rights reserved 7 Scalarm architecture Small experiment Large experiment Standard Very largeexperiment experiment © EADS 2010 – All rights reserved 8 Resource management Client Client Workload change: shorter simulations => Platform increase of management management resources overhead More workload Free resources Experiment conducting EM SiM EM SM EM EM SM SM SiM SM SiM SM SiM SiM SiM SiM SiM SiM SiM SiM Computational resources - worker node © EADS 2010 – All rights reserved EM – Experiment Manager SM – Storage Manager SiM – simulation manager 9 Scalarm applications - TODO © EADS 2010 – All rights reserved Comparison of behaviour models/soldiers strategies • Crowd behaviour depends on many input parameters: • • Behaviour model/strategy that works well for one parameters’ set may work badly for another. Example: Soldiers model created with MASDA: escalation mean: 72.22 Different strategies may be compared in 3 steps: 1. Multiple execution of mission by soldiers (using different strategies). 2. Behaviour cloning. 3. Executing data farming experiment for each cloned strategy. First implementation of soldiers model escalation mean: 153.53 Conclusion: MASDA helped to choose strategies that work well in different conditions. © EADS 2010 – All rights reserved 11 Conclusions 1. To enhance soldiers’ training, a large number of analysis of soldiers’ behaviour in different scenarios is required, thus Data Farming is a crucial module of the project. 1. EUSAS Data Farming module (implemented as Scalarm) constitutes a complete virtual platform for executing interactive Data Farming experiments. 1. Scalarm enables analysts to generate and analyze large amount of data with computer simulation in order to gain useful insight into simulated scenarios. © EADS 2010 – All rights reserved 12 Thank you for your attention ! © EADS 2010 – All rights reserved