Keynote Title: Connecting Sciences

Keynote Lecturer: Dr. Jaap van den Herik

Keynote Chair: Dr. Elias M. Awad

Presented on: 15-02-2013

Abstract: Real progress in science is dependent on novel ideas. In Artificial Intelligence, for a long time (1950-1997) a prevailing question was: can a computer chess program defeat the human World Champion? In 1997 this turned out to be the case by using a supercomputer employing the minimax method with many enhancements, such as opponent modeling. The next question was: can we use similar techniques to outperform the top grandmasters in Go. After some ten years of intensive research the answer was that minimax could not be successfully transferred to the game of Go. In the article, we will review briefly why this transfer failed. Hence, a new method for Go was developed: MCTS (Monte Carlo Tree Search). The results for MCTS in Go are rather promising. At first glance it is quite surprising that MCTS works so well. However, deeper analysis revealed the reasons for this success.
Since the field of AI-research claims to be a fruitful test bed for techniques to be applied in other areas one might look in which areas minimax or MCTS are applicable. A possible and unexpected answer is the application area of solving categories of high energy physics equations. In that area the derivation of the formulas is often performed by the (open source) computer algebra system FORM developed by Jos Vermaseren. The derivation is usually hand guided (human decisions are needed) and becomes difficult when there are many possibilities of which only a few lead to a useful solution. The intriguing question is: can we make the computer select the next step? Our idea is that the next step can be made by using MCTS.
In the article we show the first attempts which prove that this idea is realizable. It implies that games and solving high energy physics equations are connected by MCTS. A first unexpected result is showing the existence of connecting sciences. Equally unexpected is the second result. It is obtained by using MCTS for the improvement of multivariate Horner schemes when converting large formulas to code for numerical evaluation. From the viewpoint of MCTS this is an uncomplicated application and thus it is possible, by varying parameters, to see how MCTS works. It shows that the new ideas can function as cross-fertilization for two, and maybe more, research areas. Hence, the future lies in connecting sciences.

Presented at the following Conference: ICAART, International Conference on Agents and Artificial Intelligence

February, 2013
Barcelona, Spain

Conference Website:

Loading more stuff…

Hmm…it looks like things are taking a while to load. Try again?

Loading videos…