The use of machine learning to teach computers to play board games has had a lot of interest lately. Big companies such as Facebook and Google have both made recent breakthroughs in teaching AI the complex board game, Go. However, people have been using machine learning to teach computers board games since the mid-twentieth century. In the early 1960s Donald Michie, a British computer scientist who helped break the German Tunny code during the Second World War, came up with Menace (the Machine Educable Noughts And Crosses Engine). Menace uses 304 matchboxes all filled with coloured beads in order to learn to play noughts and crosses.

The National Toy Hall of Fame left a pantheon of classic playthings out of history's toy box this week when the latest inductees were announced in upstate New York. The Clue board game and the Wiffle Ball, both five-time finalists, were inducted into the hall of fame on Thursday at the Strong National Museum of Play in Rochester along with a shocker: The simple paper airplane. That brings the hall's total to 65 toys, including classics such as the Rubik's Cube, Easy-Bake Oven and Silly Putty as well as stranger picks like the stick, cardboard box and blanket. A few favorites like Rock'Em Sock'Em Robots, Transformers and Teenage Mutant Ninja Turtles have come close but failed to make the cut. Others like Trivial Pursuit, Tickle Me Elmo and backgammon have never even been finalists.

We consider the task of assigning experts from a portfolio of specialists in order to solve a set of tasks. We apply a Bayesian model which combines collaborative filtering with a feature-based description of tasks and experts to yield a general framework for managing a portfolio of experts. The model learns an embedding of tasks and problems into a latent space in which affinity is measured by the inner product. The model can be trained incrementally and can track non-stationary data, tracking potentially changing expert and task characteristics. The approach allows us to use a principled decision theoretic framework for expert selection, allowing the user to choose a utility function that best suits their objectives. The model component for taking into account the performance feedback data is pluggable, allowing flexibility. We apply the model to manage a portfolio of algorithms to solve hard combinatorial problems. This is a well studied area and we demonstrate a large improvement on the state of the art in one domain (constraint solving) and in a second domain (combinatorial auctions) created a portfolio that performed significantly better than any single algorithm.

We look at composition of (possibly nonterminating) high-level programs over situation calculus action theories. Specifically the problem we look at is as follows: given a library of available ConGolog programs and a target program not in the library, verify whether the target program executions be realized by composing fragments of the executions of the available programs; and, if so, synthesize a controller that does the composition automatically. This kind of composition problems have been investigated in the CS and AI literature, but always assuming finite states settings. Here, instead, we investigate the issue in the context of infinite domains that may go through an infinite number of states as a result of actions.  Obviously in this context the problem is undecidable. Nonetheless, by exploiting recent results in the AI literature, we devise a sound and well characterized technique to actually solve the problem.