- author = "Adrian Trenaman",
- title = "The Evolution of Autonomous Agents Using Concurrent Genetic Programming",
- school = "Department of Computer Science, National University of Ireland, Maynooth",
- year = "1999",
- address = "Ireland",
- month = oct,
- keywords = "genetic algorithms, genetic programming",
-
URL = "
http://www.cs.ucl.ac.uk/staff/W.Langdon/ftp/papers/trenaman/at_thesis1.ps.gz",
- size = "136 pages",
- abstract = "This thesis addresses the issue of how computational agents interact with and represent their environment in order to effect goal-achieving behaviour. It argues that the internal representations used by the agent to describe objects in the world should be based on how the agent perceives these objects and not necessarily on the representations a human designer might impose. A bottom-up methodology is proposed for the automatic design of distributed algorithms and internal representations to control autonomous agents. In particular, this thesis proposes and evaluates a new mechanism for the evolution of agents: {"}concurrent genetic programming''. In this encoding scheme an agent is controlled by a set of evolved programs that are executed concurrently to yield an emergent control algorithm for the agent. This encoding forms a natural interpretation of the emergent principles of the discipline of artificial life in an evolutionary context, and so elucidates the ability of evolutionary computation to create such emergent systems. The performance of the approach is investigated as a function of several parameters. These are: using different numbers of programs in the agents, explicit memory, distributed memory architectures, deterministic and non-deterministic scheduling strategies, different levels of granularity of concurrency, and the evolution of scheduling strategy. These issues are investigated through the application of concurrent genetic programming to the standard Tartarus and Dozer virtual-robotics benchmarks. It is shown that concurrent genetic programming produces better agents for these environments than a conventional genetic programming approach. It does this by employing an implicit form of state that supports the development of cyclical behaviour strategies. Implicit representations of the environment are acquired at an evolutionary level rather than at the level of the agent's experience. Although this form of internal representation leads to fit agents, it does not exhibit the formation of explicit models of the agent's environment. Instead, it allows the development of a form of internal state appropriate to achieving good fitness.",
-
notes = "
",
