Industry Session

Industrial Scheduling and Planning
Yuan Mingxuan · Huawei

Huawei is a leading global provider of information and communications technology (ICT) infrastructure and smart devices. We need to deliver hundreds of billions of dollar products across four key domains – telecom networks, IT, smart devices, and cloud services to more than 170 countries and regions on time every year. As one of the most complex supply chains in the world, a large number of scenarios rely on more efficient and intelligent scheduling & planning (S&P) algorithms. Besides supply chain, the computer system such as our storage production, circuit design and business solution supported by Huawei cloud also need to solve various S&P problems. Advanced S&P algorithms would greatly improve the efficiency of industrial systems and bring numerous economic values. In this talk, we would like to introduce some representative industrial scenarios we met as well as some of our technical works. It is expected to promote the joint research of academia and industry on advanced S&P techniques for typical practical problems.

Automated Planning and Constraint Reasoning for High Throughput Laboratory Automation
Dan Bryce · SIFT

Planning large-scale experiments for high throughput robotic cloud laboratories presents significant challenges, including choices on what measurements to take, how to allocate laboratory resources, and how to describe experiments to support execution and reproducibility. We describe our work on the DARPA Synergistic Discovery and Design project to formalize and plan high throughput screening experiments in both automated closed-loop and investigator-driven open loop experimental campaigns. We highlight the benefits to scientists gained through removing several human touch-points that lead to metadata errors and the advantages of constraint reasoning to support experiment consistency. We also describe the opportunities and challenges for planning research to address practical concerns in bringing automation to the life sciences.

Planning for Controlling Business-to-business Applications
Hector Palacios · Element AI

Automated planning is a common problem in business environments and other complex interactions between humans and organizations. However, writing a planning model is a means to an end. The popularity of deep learning cast doubts about using model-based AI as practitioners might not want to write or maintain models. The current academic and industrial discussion focus on whether learning or model-based reasoning is a better path for AI problems like policy generation. In contrast, I will discuss how the act of describing a planning model can be seen as a form of declarative control. I will comment on how simple variations of well-studied models corresponds to such control in business contexts, where policies interact with humans, organization practices, software, and data-driven agents. Focused on AI business-to-business applications, I will comment on challenges and opportunities around such new settings and recent work compatible with this direction.

Designing Goal-Oriented Conversational Agents using Automated Planning
Tathagata Chakraborti · IBM Research

Goal-oriented conversational agents, such as ones in customer support applications, require modeling of underlying business processes that end-to-end conversation models are ill-equipped to handle. In this talk, we will explore how automated planning techniques can be used in the conversational space to design large scale dialogue tree as well as scale up aggregated assistants to model sophisticated compositional behavior. We will also look at how these techniques provide out of the box explainability of agents both for the end-user as well as for the designer of these bots.

Environment Learning - Data-Driven Approaches for Real-World Decision Optimization
Wei-Wei Tu · 4Paradigm

Decision-making is the key to many real-world applications. Many decision optimization methods, e.g., reinforcement learning, have been developed for better decision making. In decision optimization tasks, we often need to interact with the real world or environment. However, accessing the real world may take risks or cost too much. For example, car accidents for automatic driving and production line adjustment in manufacturing plants for smart manufacturing are too dangerous or costly, and we can hardly afford them, and the subsequent decision making will also be badly affected. In many real-world decision optimization paradigms, we often obtain virtual environments that can mimic the real world by manually building or learning from data to avoid unnecessary risks or costs. We call the resulting research area that targets building virtual environments using experiences(e.g., data, knowledge, Etc.) gathered from real decision environments - Environment Learning. In this talk, we will go through several virtual environment driven decision optimization paradigms and present data-driven environment learning approaches that have shown significant benefits for many real-world applications, e.g., Covid-19 pandemic prevention and control.

Autonomously responding to the environment with a distributed space system
Nick Cramer · NASA

Distributed Spacecraft Systems are a type of multi-spacecraft mission architecture that can provide improved resolution, coverage, and availability of existing missions and enable missions that would be previously infeasible using traditional approaches. Autonomy is a critical need for these systems. As these systems begin to scale and have more complicated interactions, their responses to observations or operational demands need to be handled autonomously of the ground. Distributed Spacecraft Autonomy (DSA) is a project developed by the National Aeronautics and Space Administration (NASA) that enables distributed spacecraft systems by developing three capabilities: scalable communication, distributed coordination and planning, and human-swarm interaction. DSA will demonstrate these capabilities in two contexts. The first context is a flight demonstration consisting of a software payload hosted on the Starling-1 small-spacecraft mission. This software payload will use the onboard GPS receiver to perform in-situ, swarm-level reconfiguration in response to observed features in the Topside Ionosphere. The second context is a scalability study, which shows how the technologies developed in the flight demonstration can scale to a large number of spacecraft (100). The scalability demonstration applies the tools developed for the flight mission to a hardware-in-the-loop simulation of the flight software payload.