Meet the Researcher: Matt Klenk Discusses Integrated AI Applications in Transportation and Design

PARC Researcher Matt Klenk recently attended the International Conference on Automated Planning and Scheduling where he was part of an industry panel discussion on “Planning for Transportation Influence and Other Problems.” Matt was also invited to talk at the Advances in Cognitive Systems Conference at MIT, where he discussed “Cognitive Systems Research in Applications.”  We asked Matt to expand upon his recent talks and tell us more about cognitive system applications in transportation and design.

Matt, what were the common threads in these talks?

In each of these talks, and in my work more broadly, I look at the sets of models and methods that are required for intelligent behavior.

What are cognitive systems? How do they differ from traditional AI systems?

Cognitive systems are integrated AI applications that seek to capture human-level behavior on broad tasks. Many of the most pressing questions for business and society require an interdisciplinary approach to AI design. Unlike traditional AI systems, which are designed to address narrow problems and frequently only interact with domain experts, cognitive systems are typically designed using an interdisciplinary team.    

In the case of transportation, PARC works with transportation engineers, artificial intelligence researchers and user experience design experts. Likewise, we bring together different engineering disciplines, optimization experts and artificial intelligence researchers when developing AI applications in design. This interdisciplinary approach makes it easier to transition these solutions from the laboratory to practice.

What capabilities do cognitive systems enable?

The multi-scale modeling and interdisciplinary approach increases the flexibility of AI systems. By having multiple models, a system can explain its decision-making to a user, identify when its behavior does not match expectations, and diagnose which of its internal models are flawed. The right combination and organization of techniques will enable AI systems to operate in open dynamic worlds.  Currently AI companies like Alphabet employ huge teams of engineers to curate data and retrain AI systems offline whenever tasks change (e.g., search retrieving a page versus search answering a question).  The promise of cognitive systems is that it will enable continued improvement of deployed systems with minimal supervision even as tasks change. 

What are some of the limitations or challenges with cognitive systems and how do you address them?

The primary limitations of these systems are that they require significant engineering efforts with deep expertise of the AI technologies involved, along with structured background knowledge about the domain tasks. Our research addresses these challenges by understanding the principles for the design-integrated intelligent systems and by developing new AI reasoning methods to capture experience in models and reuse them in new situations.  

How are you applying cognitive systems to transportation and design?

In transportation, we developed new modeling and simulation techniques to estimate the amount of energy and time savings across Los Angeles, if we were able to send messages to 10% of the commuting population with a suggested alternative. The alternatives included switching modes, changing departure times, and even driving styles. By combining high-fidelity traffic simulations with data-driven user models, we estimated that a full deployment of our system, COPTER (Collaborative Optimization and Planning for Transportation Energy Reduction), could reduce energy consumption by 5% and congestion-induced delay by 20%.

In conceptual design, we have developed new methods for representing and reasoning over the space of all possible designs. Our graph structure captures the space of design topologies for a given number of components, and our qualitative reasoning techniques enable the analysis of sets of topologies. This allows us to search an order-of-magnitude larger space of designs than previously possible. Future design tools with this technology could assess designs that no human would dream of and, if that design is promising, provide it to the designer for consideration.

How do you envision cognitive systems impacting our world in the coming years?

I think ideas from cognitive systems that combine reasoning with learning will support new types of human-machine collaboration. This includes human-machine teams that perform a task by negotiating solutions to problems in a similar manner to traditional corporate teams. Ideas would be sketched out and promising directions would be flagged for further exploration. My vision is that eventually we will have modeling assistants that work with us and help us understand the context of our decisions.

What do you enjoy most about working at PARC?

I’m fascinated by the diversity of projects and people I have interacted with during my eight years at PARC. Due to PARC’s size, our strength is bringing together creative people from different fields to develop innovative solutions. I am constantly learning!

Furthermore, PARC straddles industry and academia in interesting ways. I publish in academic conferences and journals.  At the same time, I’m able to speak at events and meet with executives across different industries to understand how we can best solve their problems.