Creating better artificial intelligence has plenty of applications in different areas of society, from self-driving cars and aircraft to production planning, control of machines and music composition. Most current artificial intelligence research focuses on creating algorithms that can only do a single thing, or solve a single problem. To achieve artificial general intelligence we must learn how to create algorithms that can solve many different problems, without a human having to adjust the algorithm for every problem. The research in this project aims to understand how such artificial general intelligence can be created. The basic idea is to build algorithms that can create their own more specific algorithms, and learn to automatically select the right algorithm for the right problem. In order to develop these algorithms, we need a large set of good problems to test them on. Games are widely used to test AI algorithms, because they model real-world problems but are fast and easy to execute. The general algorithms developed in this project will be tested on a set of classic games, and a real-time strategy game.
This research project aims to investigate how we can create more general artificial intelligence through online stochastic search for algorithms, combined with and informed by online selection among discovered algorithms. In other words, the project will investigate the combination of genetic programming in the space of tree search algorithms with algorithm selection, also called hyper-heuristics, for creating more general problem-solving abilities. These capabilities will be evaluated through a sequence of experiments on two different test beds. Successful completion of the research will clarify the potential of search in algorithm space as a method for creating more general artificial intelligence, and produce a number of algorithms. This includes both the algorithms that will be designed for searching for algorithms and searching among algorithms, as well as the new algorithms that will be discovered by the search in algorithm space. The methods produced are expected to ultimately be generally applicable to a large number of problems.
