Massive Multi-Agent Simulation - Master Seminar

Size: px

Start display at page:

Download "Massive Multi-Agent Simulation - Master Seminar"

Avis Phillips
6 years ago
Views:

1 Massive Multi-Agent Simulation - Master Seminar Christian Hüning, BSc Hamburg University of Applied Sciences, Dept. of CS Hamburg, Germany christian.huening@haw -hamburg.de -group.org Multi Agent Research and Simulation

2 Agenda 1. A new Showcase 2. Kruger National Park Model 3. Mapping KNP to MARS 4. My Focus 5. Distribution & Scaling Concepts 6. The story so far 7. Hypotheses 8. Chances & Risks Multi Agent Research and Simulation 2

3 Maps by OpenStreetMaps MARS GROUP A new Showcase: The Kruger National Park Multi Agent Research and Simulation 3

4 Images by Thomas Thiel-Clemen MARS GROUP The Kruger National Park model km 2 of landscape Elephants Impalas 300 Cheetahs 90 meters resolution Multi Agent Research and Simulation 4

5 The Kruger National Park model Ecological Research example: Effect of water point closure on elephant density and landtype Thomas Thiel-Clemen Google Earth Multi Agent Research and Simulation 5

6 Why massive? Mengistu, D. (2011) : Many MAS applications stay behind their possibilities because of scale problems 50 % of examined MAS papers are not verifiable The two MAS papers with most agents don t even state how the results were achieved Hilbers et al. (2014) state necessity to simulate KNP on ecosystem level Cioffi-Revilla et al. (2002): Group or System size matters for systems and processes involving collective action ( ). Results need to be validated with respect to system size to ensure simulation results are not purely local for a given system size. Multi Agent Research and Simulation 6

7 Mapping the KNP model to MARS Basic MARS modelling concept Layer Approach: Borrowed from GIS data types One Layer per model aspect Composition of layers builds the environment Agents live on layers Layers are plugins T&G Coverage Layer Elevation Layer use Elephant Layer Enables: Modularization Application of Software Engineering Practices Reusability of models thanks to defined interfaces MARS - A next-gen MAS Framework 7

8 Kenvision Techniks Ltd MARS GROUP Mapping the KNP model to MARS Environment as composition of 6 different datasets: Point coverage for Elephant and Impala density Rivers Waterpoints Grass and tree coverage (T&G) Landtype Elevation Map Multi Agent Research and Simulation 8

9 Mapping the KNP model to MARS Rivers Elephants Water Points Tree & Grass Coverage Elevation Map Multi Agent Research and Simulation 9

10 Mapping the KNP model to MARS GIS data to specialized MARS GIS Layers Allows spatial queries! Useful for land type, rivers and water points Use Elevation Map to initialize Environment Service Component (ESC) Place objects in 3D space Allow efficient spatial movement and exploration Elephant and Impala models Agents Each live on their own layer Initialize elephant / impala layer with density map Multi Agent Research and Simulation 10

11 Mapping the KNP model to MARS Symbol Legend Special GIS Layer Normal Layer Elevation Layer Cheetah Layer T&G Coverage Layer Impala Layer River Layer Elephant Layer Waterhole Layer Multi Agent Research and Simulation 11

model in MARS LIFE Not a validated model from ecologists perspective

12 My Focus MARS LIFE Scalability Architecture allows for good integration and usability with MARS ecosystem Create a Kruger National Park (KNP) model in MARS LIFE Not a validated model from ecologists perspective Proof of concept for scalability of MARS LIFE Multi Agent Research and Simulation 12

13 Distribution & Scaling Concepts Model translated to MARS paradigm Distribution of model via LayerContainers in MARS LIFE, e.g.: Network LayerContainer A LayerContainer B LayerContainer C Rivers Elephants Cheetahs T&G Cover. Impalas Water Points Elevation Multi Agent Research and Simulation 13

14 Distribution & Scaling Concepts Scaling the Agents 2 distribution options: Network Network LayerContainer A LayerContainer B LayerContainer A LayerContainer B Cheetahs Impalas Cheetahs Impalas A Impalas B Agent Shadowing Multi Agent Research and Simulation 14

Distribution & Scaling Concepts Scaling the Agents Agent Shadowing: Remote-calls take a long time Every Layer instance holds all agents But: Just a small part of

15 Distribution & Scaling Concepts Scaling the Agents Agent Shadowing: Remote-calls take a long time Every Layer instance holds all agents But: Just a small part of agents as real implementation, all other entities as stubs Transparent to the developer! Caching & PUSH mechanism for agent attributes Multi Agent Research and Simulation 15

16 Distribution & Scaling Concepts Scaling the environment Environment in ecological modelling often fed from GIS files GIS files may be loaded into special LIFE layers GIS files may be splitted by attributes by features by custom filters, scripts etc. So LIFE layers may be splitted and / or replicated by the same means Read only! ESC Multi Agent Research and Simulation 16

17 Scaling the environment Network LayerContainer A LayerContainer B LayerContainer C Network LayerContainer A LayerContainer B LayerContainer C Multi Agent Research and Simulation 17

18 Distribution & Scaling Concepts Complete view: Network LayerContainer A LayerContainer B LayerContainer C Elephants Impalas A Impalas B T&G Cover. Elevation Rivers Water Points Cheetahs Elevation Multi Agent Research and Simulation 18

19 The story so far AW1 : Overview of area of research AW2 : Analysis of competing systems, concept and solutions Finding the gap (Scalability and Usability) Collecting requirements Multi Agent Research and Simulation 19

20 The story so far PO1 : Setup of experimentation environment Virtual MARS simulation cluster Controllable via WebApp Simulation As A Service! MARS Model Developer Edition to focus on model development MARS LIFE Dev Services Bare-API Service augmented simulations PO2 (In progress) : Implement missing MARS LIFE components Implement KNP model Conduct experiments & hunt bugs Multi Agent Research and Simulation 20

21 Hypotheses 1. MARS is capable to scale by a constant factor when hardware is added 2. MARS is capable of executing any given model in near real-time given enough hardware 3. It is possible to translate every model to the MARS paradigm of layers and agents 4. The MARS LIFE architecture allows to attach tools usable by domain experts whilst maintaining performance Multi Agent Research and Simulation 21

22 Chances & Risks Risks Memory consumption of Agent Shadowing might become show stopper on very(!) large models if no sufficient garbage collection algorithm can be found Chances Use massive scale models to capture Black-Swan events Allow MAS to be a tool for ecosystem scale research Enable researchers with limited hardware to work with MARS Share models, layers and agents among researchers using MARS Multi Agent Research and Simulation 22

23 Thank you for your attention! Multi Agent Research and Simulation 23

24 Sources (Hilbers et. al., 2014 ): Elephant-mediated cascading effects of water point closure in Kruger National Park, South Africa, Mengistu, D. (2011): Improving the performance of distributed multiagent based simulation. Blekinge Institute of Technology. Retrieved from Cioffi-Revilla, C. (2002): Invariance and universality in social agentbased simulations. Proceedings of the National Academy of Sciences of the United States of America, 99 Suppl 3, Multi Agent Research and Simulation 24

Interaction Design in Digital Libraries : Some critical issues

Interaction Design in Digital Libraries : Some critical issues Constantine Stephanidis Foundation for Research and Technology-Hellas (FORTH) Institute of Computer Science (ICS) Science and Technology Park