About the Workshop
The AAAI-22 workshop on AI for Decision Optimization (AI4DO) will explore how AI can be used to significantly simplify the creation of efficient production-level optimization models, thereby enabling their much wider application
and resulting business value.The desired outcome of this workshop is to drive
forward research and seed collaborations in this area by bringing together machine learning and decision-making from the lens of both dynamic and static
optimization models.
Aim and Topics
The goals of the workshop are to create joint awareness of the current state-of-the-art, including relevant software tools, and seed collaborations towards radical simplification of the creation and deployment of decision optimization models. This will be carried out through open discussions, presentations of relevant works and tools, and the exchange of ideas by researchers and practitioners working at the relevant intersection of data science, AI, and operations research, in areas such as machine learning, symbolic AI, mathematical programming, simulation and heuristic-based optimization, constraint optimization and learning, sequential decision-making and reinforcement learning, dynamic control, planning with Markov Models (MDPs, POMDPs), and learning-based optimization.
Topics of interest include, but are not limited to:
- Learning decision optimization models from data, including learning mathematical optimization
formulations (objectives and constraints), learning constraints from data, and creating reinforcement learning models from data.
- Learning dynamic models from data, including state transition models, reward models, and dynamic models for uncontrollable environmental uncertainties, with a view toward automatically
constructing environments against which agents may learn without being limited by interaction
with and/or sample complexity of real systems.
- Techniques for automated search for optimal decision-making pipelines, such as next-generation
AutoML and AutoRL: frameworks that can combine structured tabular data and expressive model-related knowledge to automatically synthesize and evaluate decision-making pipelines that are made
up of any combination of automated data transformation, automated data-driven and knowledge-driven system models, and automated model-based and data-based learning of optimal decisions.
- Techniques for accommodating model inaccuracies in data-driven learning of empirical system models, along with techniques that restrict decisions to regions of acceptable data-sufficiency, thereby
limiting unsafe extrapolation when learning empirical system models from data.
- Techniques for high-level specification of the knowledge required to create such models by users such
as domain experts, data scientists, and developers, and the derivation of such decision optimization
models from such knowledge.
- Relevant decision-focused work, such as works extending decision-focused learning to dynamic
problems, and combining decision-focused machine learning with classical machine learning to
handle objective functions and system constraints that are learned empirically from historical data.
- Ensuring the end-to-end accuracy and performance of data- and knowledge-based generated models,
by accounting for the limitations of the data and knowledge used to generate them.
- Techniques for speeding up the solution time of such generated models, including specification of
knowledge that can be used to speed up the solution time, and using AI to accelerate the solution
of such automatically generated models (including combinatorial optimization problems), to ensure
that these models can be solved in an acceptable amount of time.
Program
February 28, 2022. All times are in PST (Pacific Coast US).
|
06:45 - 07:40 AM
|
Introduction and Keynote
|
|
|
AI for Engineering
Pascal Van Hentenryck,
the A. Russell Chandler III Chair and Professor, and the associate chair of innovation and entrepreneurship in the H. Milton Steward School of Industrial and Systems Engineering at the Georgia Institute of Technology
Abstract: This talk argues that the fusion of machine learning and optimization
is a promising avenue to address fundamental challenges in engineering
domains, including in power systems, logistics and supply chains,
mobility, and circuit design. The presentation presents various
case studies where optimization proxies, learning to optimize, and
end-to-end learning are pushing the frontiers in these domains.
|
|
07:40 - 07:45 AM
|
Break
|
|
07:45 - 08:30 AM
|
Lightning Demo Session
|
|
|
07:45 AM
Knowledge-driven Decision Optimization for Non-Experts [article]
Yishai A. Feldman, Aviad Sela and Segev Wasserkrug
08:00 AM
OptiCL: A Package for Mixed-Integer Optimization with Constraint Learning [article]
Donato Maragno, Holly Wiberg, Dimitris Bertsimas, S. Ilker Birbil, Dick den Hertog and Adejuyigbe Fajemisin
08:15 AM
Automated Decision Optimization: Data and Knowledge-driven Optimization Model Generation with Human-in-the-loop
Lisa Amini, Arunima Chaudhary, Yishai Feldman, Pavankumar Murali, Lam Nguyen, Dzung Phan, Aviad Sela, Carolina Spina, Dharmashankar Subramanian, Abel Valente, Long Vu, Dakuo Wang, Segev Wasserkrug, Ritesh Yadav and Nianjun Zhou
|
|
08:30 - 08:45 AM
|
Break
|
|
08:45 - 10:00 AM
|
Paper Session
|
|
|
08:45 AM
Almost Hyperparameter-free Hyperparameter Selection Framework for Offline Policy Evaluation[article]
Kohei Miyaguchi
09:00 AM
A Prediction-based Approach for Online Dynamic Radiotherapy Scheduling [article]
Tu-San Pham, Antoine Legrain and Louis-Martin Rousseau
09:15 AM
GDI: Rethinking What Makes Reinforcement Learning Different from Supervised Learning
Jiajun Fan, Changnan Xiao and Yue Huang
09:30 AM
Logical Neural Networks to Serve Decision Making with Meaning
Lan Hoang and Alexander Zadorojniy
09:45 AM
Automated Decision Optimization with Reinforcement Learning [article]
Radu Marinescu, Tejaswini Pedapati, Long Vu, Paulito Palmes, Todd Mummert, Peter Kirchner, Dharmashankar Subramanian, Parikshit Ram and Djallel Bouneffouf
|
|
10:00 - 10:30 AM
|
Long Break
|
|
10:30 - 11:00 AM
|
Special Talk
|
|
|
Optimization with Constraint Learning: A Framework and Survey Extended Abstract [article]
Adejuyigbe Fajemisin, Donato Maragno and Dick den Hertog
|
|
11:00 AM - 12:00 PM
|
Panel Driven Brainstorming Session
|
|
|
The Future of AI for Decision Optimization
Panel Members:
Prof. David Bergman, University of Connecticut
Prof. Bistra Dilkina, USC
Prof. Pascal Van Hentenryck, Georgia Institute of Technology
Prof. Frank Hutter, Freiburg University
Prof. Michele Lombardi, University of Bologna
Dr. Segev Wasserkrug, IBM Research
Holy Wiberg, MIT
|
|
12:00 - 12:30 PM
|
Break and Breakout Discussions / Networking
|
|
12:30 - 01:45 PM
|
Paper Session
|
|
|
12:30 PM
Predict and Optimize: Through the Lens of Learning to Rank [article]
Jayanta Mandi, Victor Bucarey Lopez, Maxime Mulamba Ke Tchomba and Tias Guns
12:45 PM
Addressing Solution Quality in Data-generated Optimization Models [article]
Orit Davidovich, Parikshit Ram, Segev Wasserkrug, Dharmashankar Subramanian, Nianjun Zhou, Dzung Phan, Pavankumar Murali and Lam Nguyen
01:00 PM
Constraint Acquisition and the Data Collection Bottleneck [article]
Steve Prestwich
01:15 PM
End-to-End Learning via Constraint-Enforcing Approximators for Linear Programs with Applications to Supply Chains [article]
Rares Cristian, Pavithra Harsha, Georgia Perakis, Brian Quanz and Ioannis Spantidakis
01:30 PM
JUMBO: Scalable Multi-task Bayesian Optimization Using Offline Data [article]
Kourosh Hakhamaneshi, Pieter Abbeel, Vladimir Stojanovic and Aditya Grover
|
|
01:45 - 02:00 PM
|
Summary and Next Steps Discussion
|
|
02:00 PM
|
End of Workshop
|
Call for workshop papers
Authors are invited to send contributions of both relevant works and relevant software tools, both in
the AAAI-22 proceedings format (see AAAI author kit). Submission is through the AAAI-22 AI4DO
Workshop Easychair Submission Link.
Contributions of relevant work should be in one of the following forms:
- Long paper: Submission of original work up to eight pages in length (including references).
- Short paper: Submission of work in progress with preliminary results, and position papers, up to four pages in length (+ references).
- Extended abstract: Summary of recently published journal/conference papers in the form of a
two-pages extended abstracts.
- Fast Track (Rejected AAAI papers): AAAI papers that were rejected in the second round of reviews, with *average* scores of at least 5.0 may be submitted, along with previous reviews and scores and an optional letter indicating how the authors have addressed the reviewers' comments.
Software tool submissions should include:
- A two-page short paper (+ one-page reference) describing the tool, which should include the technical details of the tool and related work, and describe the tool's significance and relevance to the workshop
topics.
- A demonstration video (or publicly available link of such video) of the tool, or link to a
relevant website (which could also be a github link) from where the tool can be downloaded.
The review process is single-blind. The program committee will select the papers and tools to be
presented at the workshop according to their suitability to the aims.
