Design is the Ultimate AI Challenge

This article was written by PARC Researchers, Johan de Kleer and Sai Nelaturi.

Johan leads the Model-based Reasoning area in the System Sciences Laboratory at PARC. Sai manages the Computation for Automation in Systems Engineering area in the System Sciences Laboratory at PARC.

By far most of the AI-directed effort is focused on making our personal assistants, weapon systems, cities, … behave as intelligently as possible. We can embed as much intelligence in the systems of our modern civilization as we want, but the ultimate capability of the overall system will always be limited by its design. The hallmark of our humanness is our creativity. Almost all current AI research misses the aspects of intelligence that are needed for creative design. This is profoundly unfortunate, because the foundations for work on creative design now exist.

Design is the process of converting function to form. It is an activity distinct from reasoning, learning and the like (although it builds on them) in that design involves simultaneous planning for an artifact’s function in terms of its form, materials, and manufacture. A designed artifact has a life of its own apart from the designer. Designing for a-priori unknown functions is particularly challenging. Thus, a designer needs to plan for unknown future uses. The range of pieces from which we compose and create our designs is growing exponentially. Just think of the new materials and their combinations we hear about every day. 3D printing, hybrid manufacturing, and other advanced fabrication techniques are also enabling unprecedented complexity. We are in this very strange world where our ability to manufacture an artifact far outstrips our ability to design them.

Design invariably requires the combination of humans and computers working together. There is no equivalent to the even limited Google ability to answer questions for design. No design tool today in any domain comes remotely close to what an expert designer can do. You might ask, what does AI have to contribute to design thinking? By far the most common design automation approach is searching enormous design spaces (invariably some kind of optimization; Topology Optimization being a very popular technique). This is not how people creatively design. Combinatorial search or machine learning by themselves are incapable of explaining the design process to humans. Moreover, they scale very poorly. Most of the important design concepts come from humans. Engineering design practice is critical, but progress in AI technologies for design is key for the 21st century:

• Use of analogy throughout design. Most designs are based on concepts from other designs
• Explicit (vs. usually implicit) representation of design knowledge, and common sense physical reasoning
• Simultaneous reasoning at multiple levels of abstraction
• Exploring otherwise gigantic spaces efficiently exploiting mathematical properties and re-representation
• Deep integration of physics and data driven approaches with model-based reasoning

Here at PARC, we are exploring this integration of AI and engineering thinking that is so vital to creative design.