Research Interests

• Design Theory & Methodology
• Cognitive Modeling & Artificial Intelligence in Design
• Learning language and concepts in Design
• Computer Aided Design
• Robotics, Motion Planning
• Optimization

 

Symbols in Design

Dimensionality Reduction in Design

Unsupervised Learning in Design

 

 

Research Summary

Starting as a baby: How to evolve CAD systems that learn from experience

Today's CAD systems are unable to abstract knowledge via explorations in the design space, leading to an inability to generalize across similar designs, so that both structural chunks (symbols) as well as behavioral patterns (rules) need to be programmed by users and cannot be learned. If such a system could be created, it may be thought of as a baby designer, who can learn from exploring problem spaces, but who does not have the ability to actually design. In this work, we take the position that in order to eventually build systems that can generalize and learn about their design tasks, we must first build such a baby designer, and then expose it to increasing amounts of experience to build the knowledge that eventually is needed in executing design tasks. Although such approaches have been used widely in AI and natural language processing , the approaches to intelligent design have largely been rule-based or case-based, which are known to be susceptible to brittleness. Here we attempt to provide a very elementary computational approach that begins to address some of the issues related to creating a baby designer. Instead of attempting to pre-program the domain knowledge relevant to a design task, we attempt to acquire this knowledge as patterns of high performance within well-explored design spaces, and explore the possibility of transferring some of the learned patterns for other designs in similar part families. These patterns (of feasible designs) often result in significant reductions in the dimensionality of the design space, and may reveal structural chunks that appear in different tasks and therefore may eventually be recognized as what designers today use as symbols. This work constitutes only the first stage in evolving such a baby designer, and we demonstrate some elementary capabilities for this system: a) Emergence, or the discovery of patterns in the design space b) How the emergent patterns change under design change, i.e. for a new embodiment or with additional design variables, and c) How the function model can be used to obtain unsupervised clusters that may constitute a first cut towards product portfolio optimization. more :