Artificial Intelligence: An Approach

The ongoing and seemingly unstoppable digital transformation brings forth new options, opportunities but also challenges to individuals, organizations, companies and societies alike. Governments are alarmed, realizing the potential consequences on the workforce, while also being apparently helpless against uncontrollable and powerful digital players such as Google or Facebook. As Henry Wittke shows, recent breakthroughs in the field of machine learning increase the potential of Artificial Intelligence to disrupt the world's largest industries.

I am indebted to my thesis advisor, Michael Genesereth, for his guidance, inspiration, and support which has made this research possible. As a teacher and a sounding board for new ideas, Mike was extremely helpful in pointing out Haws, and suggesting new directions to explore. I would also like to thank Harold Brown for introducing me to the application of artificial intelligence to reasoning about designs, and his many valuable comments as a reader of this thesis. Significant contribu­ tions by the other members of my reading committee, Mark Horowitz, and Allen Peterson have greatly improved the content and organization of this thesis by forcing me to communicate my ideas more clearly. I am extremely grateful to the other members of the Logic Group at the Heuristic Programming Project for being a sounding board for my ideas, and providing useful comments. In particular, I would like to thank Matt Ginsberg, Vineet Singh, Devika Subramanian, Richard Trietel, Dave Smith, Jock Mackinlay, and Glenn Kramer for their pointed criticisms. This research was supported by Schlumberger Palo Alto Research (previously Fairchild Laboratory for Artificial Intelligence). I am grateful to Peter Hart, the former head of the AI lab, and his successor Marty Tenenbaum for providing an excellent environment for performing this research.

Routing of VLSI chips is an important, time consuming, and difficult problem. The difficulty of the problem is attributed to the large number of often conflicting factors that affect the routing quality. Traditional techniques have approached routing by ignoring some of these factors and imposing unnecessary constraints in order to make routing tractable. In addition to the imposition of these restrictions, which simplify the problems to a degree but at the same time reduce the routing quality, traditional approaches use brute force. They often transform the problem into mathematical or graph problems and completely ignore the specific knowledge about the routing task that can greatly help the solution. This thesis overcomes some of the above problems and presents a system that performs routing close to what human designers do. In other words it heavily capitalizes on the knowledge of human expertise in this area, it does not impose unnecessary constraints, it considers all the different factors that affect the routing quality, and most importantly it allows constant user interaction throughout the routing process. To achieve the above, this thesis presents background about some representative techniques for routing and summarizes their characteristics. It then studies in detail the different factors (such as minimum area, number of vias, wire length, etc.) that affect the routing quality, and the different criteria (such as vertical/horizontal constraint graph, merging, minimal rectilinear Steiner tree, etc.) that can be used to optimize these factors.

Why yet another book on Artificial Intelligence?

In 1984, while I was following his postgraduate course on VLSI design, my supervisor Dr. David Skellern, asked me if I was interested in investigating intelligent approaches to automatic Ie floorplanning. He told me then: "a circuit that works always looks nice, has a clever data and control flow.

This book introduces the novel artificial intelligence technique of polymodels and applies it to the prediction of stock returns. The idea of polymodels is to describe a system by its sensitivities to an environment, and to monitor it, imitating what a natural brain does spontaneously. In practice this involves running a collection of non-linear univariate models. This very powerful standalone technique has several advantages over traditional multivariate regressions. With its easy to interpret results, this method provides an ideal preliminary step towards the traditional neural network approach.