Computational Science Technical Note CSTN-088

CSTN Home

Architecture design for self-training intelligent vehicle-driving agents: paradigms and tools

A. P. Gerdelan

Archived April 2009

Abstract

Enabling vehicle-driving intelligent agents to improve their own driving behaviour requires a specialised simulation framework. The design of such a framework is no trivial task, and a number of critical design questions are raised. Numerous training machine-learning paradigms are available; brute-force methods, evolutionary selection techniques and genetic algorithms - what are the training/performance trade-offs of each? How do we choose from the wide variety of variables that can be used to punish or reward the agents? There is also a completely open design choice when it comes to choosing a training arena - should the agents learn during application or in pre-built training courses? We contrast and compare a range of different methods for application to our particular agents, evaluate the use of different levels of complexity for fitness evaluation functions, and explore differences in framework design to optimise either navigation performance or human-perceived realism of simulated motion through machine learning techniques, and discuss some new tools that we have created to aid our design.

Keywords: machine-learning; 3D graphics; intelligent agents; virtual reality; video game AI; fuzzy logic.

Full Document Text: PDF version.

Citation Information: BiBTeX database for CSTN Notes.

BiBTeX reference:

@TECHREPORT{CSTN-088,
  author = {A. P. Gerdelan},
  title = {Architecture design for self-training intelligent vehicle-driving
	agents: paradigms and tools},
  institution = {Computer Science, Massey University},
  year = {2009},
  number = {CSTN-088},
  address = {Albany, North Shore 102-904, Auckland, New Zealand},
  month = {April},
  timestamp = {2009.09.06}
}


[ CSTN Index | CSTN BiBTeX ]