Designing AI models of natural intelligence isn’t as easy as it should be. My research therefore has always included a great deal of work on systems AI (intelligence by design), including action selection and the development methodology I initially developed at MIT, Behavior Oriented Design (BOD). We apply this work in a variety of domains besides science, including cognitive robotics, computer game characters, and “smart homes.”
Designing Intelligent Systems
All artificial systems must be created, and as such they need good design. Learning, genetic algorithms and autonomous development are useful techniques, but they don't spontaneously create working artificial systems. In nature, evolution took billions of years to create the genetic scaffolding essential for animals with complex intelligence. Natural human-like intelligence further requires many years of individual and social learning. Artificial systems have to be designed and constructed before they can begin to do any learning, evolving or developing, or perform any other function.
Behavior Oriented Design (BOD) is a development methodology for intelligent and cognitive systems. The BOD methodology has been used to develop software for robots, real-time virtual reality characters, intelligent environments, and experimental platforms for increasing our understanding of natural intelligence. I have also done some work on applying this methodology to designing services for the semantic web, creating intelligent tutoring systems and managing ubiquitous computing in intelligent environments. These latter projects have not yet been fully developed, but I mention them here because they give a broader notion of what an intelligent system can be.
BOD is a modular approach that extends object-oriented design to support agency. Basic actions are provided by a reusable modular library developed for a particular platform and domain (e.g. a set of domestic robots, or characters for a virtual game world). The process of specifying agency can also be seen as the process of system integration around a particular individual intelligence. Both are achieved by specifying the agent’s priorities using POSH dynamic plans. These plans facilitate intelligent action selection by providing priorised arbitration wherever behaviour generated by two or more different modules might otherwise be in conflict.
Action Selection
Action selection is the means by which an agent (either an animal or an autonomous artificial system) determines at any instant what to do next. For AI developers, action selection is also a key mechanism for integrating the design of intelligent systems. The term action selection does not imply any conscious or deliberate choice, but is rather a functional description of the process of generating intelligent behaviour.
There are two key questions in action selection:
What is being selected?
How is it being selected?
Theories of action selection range from completely dynamic models, where there are never any discrete acts being selected but only continuous integrated processes resulting in emergent behaviour, to logic-based strictly-sequential provably-optimal lists of actions referred to as plans. In natural intelligence, we know that some action selection is performed in a distributed manner. For example, some actions are controlled from the spine independently of the brain. But we also know that complex discrete actions are represented by and can be generated from the activation of single nerve cells.
How action selection works in nature is a core research question for the Artificial Models of Natural Intelligence (AmonI) group at Bath, while producing AI action selection is one of our core technologies.