- author = "David Browne",
- title = "Vision-Based Obstacle Avoidance: A Coevolutionary Approach",
- school = "Department of Software Development, Monash University",
- year = "1996",
- type = "Bachelor of Computing with Honours",
- address = "Australia",
- month = oct,
- keywords = "genetic algorithms, genetic programming",
-
URL = "
http://www.csse.monash.edu.au/hons/projects/1996/David.Browne/",
-
URL = "
http://www.cs.ucl.ac.uk/staff/W.Langdon/ftp/papers/browne/browne_thesis.ps.gz",
- size = "147 pages",
-
abstract = "This thesis investigates the design of robust obstacle
avoidance strategies. Specifically, simulated
coevolution is used to breed steering agents and
obstacle courses in a `computational arms race'. Both
steering agent strategies and obstacle courses are
represented by computer programs, and are coevolved
according to the genetic programming paradigm.
Previous research has found it difficult to evolve robust vision based obstacle avoidance agents. By independently evolving obstacle avoidance agents against a competing evolving species (ie the obstacle courses), it is hypothesised that the robustness of the agents will be increased.
The simon system, an existing genetic programming tool, is modified and used to evolve both the obstacle avoidance agents and the obstacle courses. A comparison is made between the robustness of coevolved obstacle avoidance agents and traditionally evolved (non-coevolved) agents. Robustness is measured by average performance in a series of randomly generated obstacle courses.
Experimental results show that the average robustness of the coevolved oa agents is greater than that of the traditionally evolved, and statistically it is shown that this data is representative of all cases.
It is therefore concluded that coevolution is applicable to oa type problems, and can be used to evolve more robust, general purpose Vision-Based Obstacle Avoidance agents.",
