Presentation of a model designed to easily generate various and consistent agents' behaviors in simulations. The model is applied to traffic simulation in Renault's driving simulators.
This is the presentation of the paper entitled "Generating Various and Consistent Behaviors in Simulations" at the 2009 International Conference on Practical Applications of Agents and Multi-Agent Systems (PAAMS'09).
Generating Various and Consistent Behaviors in Simulations
1. Generating Various and Consistent Behaviors
in Simulations
Benoît Lacroix 1,2, Philippe Mathieu 2 and Andras Kemeny 1
1 Renault, Technical Center for Simulation
2 LIFL, University of Lille March 26, 2009 PAAMS 2009
2. Context and motivation
Renault / LIFL UMR CNRS collaboration
Context: traffic simulation in driving simulators
Evaluation of ergonomics, embedded systems, design…
Needs
Various and consistent behaviors for autonomous vehicles (cautious, aggressive…)
Usable by scenario designers
Idea
Driving psychologists classify drivers depending on their behavior (Saad, 1992)
Drivers use set of norms (based on Highway Code, informal rules…)
But they do not strictly follow these norms
Generic approach to address the issue
Behaviors description using norms
Generation engine managing the determinism
Monitoring
Benoit Lacroix
Renault, Technical Center for Simulation March 26, 2009 PAAMS 2009 2
3. Normative description of behaviors
Normative systems (Noriega, 1997; Esteva et al., 2001; Vazquez-Salceda et al., 2005)
Organizational control in multi-agent based simulations
Improve agents coordination, communication…
In our case
Institution: parameters and associated definition domains
Norms: subsets of these parameters and domains
Behaviors: instantiations of these norms
For instance, in traffic
Parameters: maximal speed, safety time…
Institution: bounds of these parameters (max speed in [0,300] km/h)
Norms: cautious, aggressive drivers (max speed in [140,160] km/h)
Behavior: a cautious, an aggressive (max speed = 156 km/h)
Benoit Lacroix
Renault, Technical Center for Simulation March 26, 2009 PAAMS 2009 3
4. Generation engine
Variety
Randomly select parameters
from a norm
Behavioral variety within a
norm
Allow violations: one or more
parameters outside the limits
Consistency
Guaranteed when generation within norms limits
Mechanism to reject aberrant behaviors (quantification)
Reaction to violations at runtime
Benoit Lacroix
Renault, Technical Center for Simulation March 26, 2009 PAAMS 2009 4
5. Monitoring
Emergence of new norms
Feedback to the users
Improve design and calibration
Calibration with real data
Learning norms from real data sets
Unsupervised learning
Kohonen Neural Networks
Description of the data space
Linear component analysis
Benoit Lacroix
Renault, Technical Center for Simulation March 26, 2009 PAAMS 2009 5
6. Application
Application
Driving simulation software SCANeR™ II
Ergonomics, embedded systems, design, headlights…
Description
Agents’ decision model: perception – decision (finite state automata) –
action (vehicle dynamic model)
Institution parameters = existing vehicles parameters of traffic model
Traffic managed by the existing model
Uses
Introduction of driving styles
Generation of the “ambient” traffic
Benoit Lacroix
Renault, Technical Center for Simulation March 26, 2009 PAAMS 2009 6
8. Experimental results
Highway database
11 km, 3000 veh/h
Normal, aggressive and
cautious drivers
Speed distributions
More norms increase variety
Increased dynamicity
Lane repartition
Aggressive on left lane
Cautious on right lane
Benoit Lacroix
Renault, Technical Center for Simulation March 26, 2009 PAAMS 2009 8
9. Conclusion
Easily create various behaviors
Manage the generation process
Guaranty the consistency of the behaviors
Allow violations if wished
Wide application range
Non-intrusive
Perspectives
Norms calibration with real data
Benoit Lacroix
Renault, Technical Center for Simulation March 26, 2009 PAAMS 2009 9
10. Thank you for your attention
Contact: benoit.lacroix@gmail.com
Benoit Lacroix
Renault, Technical Center for Simulation March 26, 2009 PAAMS 2009 10