MSc. Project Presentation: Statistical modeling and analysis of autonomous agents' decisions in learning to cross cellular automaton based highway

Date and Time

Location

Summerlee Science Complex Room 1504

Details

CANDIDATE:  Chong Gan

ABSTRACT:

In this project, statistical analysis of simulation model of autonomous agents’ decisions in learning to cross cellular automaton based highway is conducted. The objective for autonomous agents is to cross the highway successfully, which leads to four different types of decisions that is correct crossing decisions, incorrect crossing decisions, correct waiting decisions, and incorrect waiting decisions. The statistical analysis is to explore the main component in the configuration parameters of the model and to investigate how this main factor influences the agents’ decision-making ability. Particularly, since the parameter knowledge-based transfer (i.e., transfer of the knowledge base based on agents’ prior crossing experience) is a key for agents to determine the self-learning method, we constantly focus on the role that it plays through different analytical techniques. The other purpose of the project is to reduce the complexity of the simulation data by looking for a way to simplify the analyzed data set. Thus, we utilize (1) canonical correlation analysis and regression tree analysis to examine the relationship between decisions and parameters, (2) linear regression to discuss how the KB transfer can play a leading role in the agents’ decision-making abilities in different traffic density, (3) principal component analysis to reduce the dimension of the data frame while preserving the main effects of the relationships, (4) k-means clustering and ANOVA F-test to identify the simulation setups that share the commonalities, (5) two novel approaches (i.e., smoothing based gap statistic and maximum ratio of qualified clusters) to determine the optimal number of clusters for k-means clustering. The results based on various statistical analysis provide a comprehensive understanding of the agents’ learning process and recommend the improvement of the simulation.

 

Advisory Committee

  • A. Lawniczak, Advisor
  • G. Umphrey
  • S. Xie

Find related events by keyword

Events Archive