Evolutionary Networks for Multi-Behavioural Robot Control : A thesis presented in partial fulfilment of the requirements for the degree of Master of Science in Computer Science Massey University, Albany, New Zealand

Abstract

Artificial Intelligence can be applied to a wide variety of real world problems, with varying levels of complexity; nonetheless, real world problems often demand for capabilities that are difficult, if not impossible to achieve using a single Artificial Intelligence algorithm. This challenge gave rise to the development of hybrid systems that put together a combination of complementary algorithms. Hybrid approaches come at a cost however, as they introduce additional complications for the developer, such as how the algorithms should interact and when the independent algorithms should be executed. This research introduces a new algorithm called Cascading Genetic Network Programming (CGNP), which contains significant changes to the original Genetic Network Programming. This new algorithm has the facility to include any Artificial Intelligence algorithm into its directed graph network, as either a judgement or processing node. CGNP introduces a novel ability for a scalable multiple layer network, of independent instances of the CGNP algorithm itself. This facilitates problem subdivision, independent optimisation of these underlying layers and the ability to develop varying levels of complexity, from individual motor control to high level dynamic role allocation systems. Mechanisms are incorporated to prevent the child networks from executing beyond their requirement, allowing the parent to maintain control. The ability to optimise any data within each node is added, allowing for general purpose node development and therefore allowing node reuse in a wide variety of applications without modification. The abilities of the Cascaded Genetic Network Programming algorithm are demonstrated and proved through the development of a multi-behavioural robot soccer goal keeper, as a testbed where an individual Artificial Intelligence system may not be sufficient. The overall role is subdivided into three components and individually optimised which allow the robot to pursue a target object or location, rotate towards a target and provide basic functionality for defending a goal. These three components are then used in a higher level network as independent nodes, to solve the overall multi- behavioural goal keeper. Experiments show that the resulting controller defends the goal with a success rate of 91%, after 12 hours training using a population of 400 and 60 generations.