Genetic Context Learning Automatic behavior modeling from observed human performance

1
Genetic Context Learning Automatic behavior
modeling from observed human performance

• Hans Fernlund, Sven Eklund and Avelino Gonzalez

2
Problem Statement
Automatically Create
Simulated Agents with
individual Human Behavior
Human Behavior
by Observation
Context-Based Reasoning (CxBR)
Genetic Context Learning (GenCL)
Genetic Programming (GP)
3
CxBR GP gt GenCL

• Context-Based Reasoning (CxBR)
• Situational Awareness
• Hierarchical structure
• Limits Search Space
• Intuitive

• Genetic Programming (GP)
• Evolutionary Algorithm
• Applicable to many problem domains
• Non transforming

Generation
i1
Generation i
Genetic
Evaluation

Selection

operations





Iterate until end condition




4
Experiments Data

• Evolve simulated cars with human characteristics
• Commercial Driving Simulator
• Five drivers gt five different agents
• Two data sets training and validation
  environment
• Learning by Observation
• City driving
• Realistic Environment
• No repeated situations
• Unpredictable behavior

5
Data set and Context base

• Data set reduced
• lt 350 of 5000 samples used per agent
• Pre-defined empty context structure

Urban-Driving
Intersection-Turning
Traffic-Light-Driving
Red-Light-Driving
Green-Light-Driving
6
Evaluation Criteria

• Learning capabilities
• Error rate on the training data
• Generalization
• Capable of handling new situations
• Long term reliability
• Autonomous agents long term performance
• Usefulness
• Effectiveness compared to traditional techniques

7
Conclusions and Results

• Features of GenCL agents
• Learn, generalize
• Reliable agents
• Individual behavior patterns
• Competitive with human modeling performance
• Learning in all context parts in CxBR
• Transparent knowledge usable for interpretation
• Capable to learn by observation

8
Thank you for your attention

• Questions?

9
Generalization in the validation environment

• Qualitative validation
• Quantitative validation

10
Long term reliability

• 40 minutes of simulation time
• gt60 traffic lights (time triggered)
• All agents running after 40 minutes
  intersection turning consistency

11
Usefulness

• Comparison to agent developed by Knowledge
  Engineer
• Training environment
• Validation environment