Abstract

Much research on intelligent systems has concentrated on low level mechanisms or sub-systems of restricted functionality. We need to understand how to put all the pieces together in an *architecture* for a complete agent with its own mind, driven by its own desires. A mind is a self-modifying control system, with a hierarchy of levels of control, and a different hierarchy of levels of implementation. AI needs to explore alternative control architectures and their implications for human, animal, and artificial minds. Only within the framework of a theory of actual and possible architectures can we solve old problems about the concept of mind and causal roles of desires, beliefs, intentions, etc. The high level ``virtual machine'' architecture is more useful for this than detailed mechanisms. E.g. the difference between connectionist and symbolic implementations is of relatively minor importance. A good theory provides both explanations and a framework for systematically generating concepts of possible states and processes. Lacking this, philosophers cannot provide good analyses of concepts, psychologists and biologists cannot specify what they are trying to explain or explain it, and psychotherapists and educationalists are left groping with ill-understood problems. The paper sketches some requirements for such architectures, and analyses an idea shared between engineers and philosophers: the concept of ``semantic information''.