This talk explores the potential of using symbolic machine learning systems to provide adaptivity in agents that are based on logical foundations, such as BDI agents. Inspiration for this approach comes from seminal work of Dzeroski et al on relational reinforcement learning, in which the TILDE decision-tree learning system was employed to approximate the policy and Q-function in various experiments in blocks world. In the approach to be discussed, the symbolic learning system is Alkemy, a decision-tree learning system with a foundation based on higher-order logic. The key idea is to provide adaptivity by an on-line learning process with the current policy as an Alkemy decision-tree, the space of possible policies as an Alkemy hypothesis language, and a stream of training examples to guide the changes in the policy. An attractive aspect of this approach is that the policies are symbolic and therefore comprehensible. Thus initial policies can be directly written by agent designers and users of the agents can be given comprehensible explanations of their behaviour. The main components of the architecture will be illustrated by a simple example. Along the way, a setting for predicate construction in higher-order logic used by Alkemy will be presented.
The research to be described is being carried out in the context of the Smart Internet Technology Cooperative Research Centre, a substantial 7 year Australian research initiative having the overall research goal of making interactions that people have with the Internet much simpler than they are now. One of the research programs in the CRC is concerned with building Smart Personal Assistants and one project in that program is concerned with building adaptive agents, the topic of this talk.