ICAPSICAPS 2024 Keynote: Dale Schuurmans
Professor, University of Alberta Research Scientist, Google DeepMind
Computing and Planning with Large Generative Models
The ability of large generative models to respond naturally to text, image and audio inputs has created significant excitement. Particularly interesting is the ability of such models to generate outputs that resemble coherent reasoning and computational sequences. I will first discuss the inherent computational properties of large language models, showing how they can be proved Turing complete in natural deployments. The co-existence of informal and formal computational systems in the same model does not change what is computable, but does provide new means for eliciting desired behaviour. I will then consider non-deterministic computation, which captures planning and theorem proving as special cases. Finally, I will discuss some recent progress in leveraging large text-video models as real world simulators that enable planning for real environments. Leveraging large generative models jointly as simulators and agents has led to advances in several application areas.
Short Bio
Dale Schuurmans is a Research Director at Google DeepMind, Professor of Computing Science at the University of Alberta, a Canada CIFAR AI Chair, and a Fellow of AAAI. He has served as an Associate Editor-in-Chief for IEEE TPAMI, an Associate Editor for JMLR, AIJ, JAIR and MLJ, and as a Program Co-chair for AAAI-2016, NeurIPS-2008 and ICML-2004. He has published over 250 papers in machine learning and artificial intelligence, and received paper awards at NeurIPS, ICML, IJCAI, and AAAI.
ICAPS 2024 Keynote: Dale Schuurmans on Computing and Planning with Large Generative ModelsICAPS2024-07-02 | ICAPS 2024 Keynote: Dale Schuurmans
Professor, University of Alberta Research Scientist, Google DeepMind
Computing and Planning with Large Generative Models
The ability of large generative models to respond naturally to text, image and audio inputs has created significant excitement. Particularly interesting is the ability of such models to generate outputs that resemble coherent reasoning and computational sequences. I will first discuss the inherent computational properties of large language models, showing how they can be proved Turing complete in natural deployments. The co-existence of informal and formal computational systems in the same model does not change what is computable, but does provide new means for eliciting desired behaviour. I will then consider non-deterministic computation, which captures planning and theorem proving as special cases. Finally, I will discuss some recent progress in leveraging large text-video models as real world simulators that enable planning for real environments. Leveraging large generative models jointly as simulators and agents has led to advances in several application areas.
Short Bio
Dale Schuurmans is a Research Director at Google DeepMind, Professor of Computing Science at the University of Alberta, a Canada CIFAR AI Chair, and a Fellow of AAAI. He has served as an Associate Editor-in-Chief for IEEE TPAMI, an Associate Editor for JMLR, AIJ, JAIR and MLJ, and as a Program Co-chair for AAAI-2016, NeurIPS-2008 and ICML-2004. He has published over 250 papers in machine learning and artificial intelligence, and received paper awards at NeurIPS, ICML, IJCAI, and AAAI.ICAPS 2024 Keynote: Julie ShahICAPS2024-06-28 | ICAPS 2024 Keynote: Julie Shah
Effective Human-Machine Partnerships in High Stakes Settings
Every team has top performers -- people who excel at working in a team to find the right solutions in complex, difficult situations. These top performers include nurses who run hospital floors, emergency response teams, air traffic controllers, and factory line supervisors. While they may outperform the most sophisticated optimization and scheduling algorithms, they cannot often tell us how they do it. Similarly, even when a machine can do the job better than most of us, it can’t explain how. The result is often an either/or choice between human and machine - resulting in what we call zero-sum automation. In this talk I present research case studies from industry and also share our lab's latest research effectively blending the unique decision-making strengths of humans and intelligent machines.
Short Bio
Julie Shah is the H.N. Slater Professor and Head of Aeronautics and Astronautics, faculty director of MIT's Industrial Performance Center, and director of the Interactive Robotics Group, which aims to imagine the future of work by designing collaborative robot teammates that enhance human capability. She is expanding the use of human cognitive models for artificial intelligence and has translated her work to manufacturing assembly lines, healthcare applications, transportation and defense. Before joining the faculty, she worked at Boeing Research and Technology on robotics applications for aerospace manufacturing. Prof. Shah has been recognized by the National Science Foundation with a Faculty Early Career Development (CAREER) award and by MIT Technology Review on its 35 Innovators Under 35 list. She was also the recipient of the 2018 IEEE RAS Academic Early Career Award for contributions to human-robot collaboration and transition of results to real world application. She has received international recognition in the form of best paper awards and nominations from the ACM/IEEE International Conference on Human-Robot Interaction, the American Institute of Aeronautics and Astronautics, the Human Factors and Ergonomics Society, the International Conference on Automated Planning and Scheduling, and the International Symposium on Robotics. She earned degrees in aeronautics and astronautics and in autonomous systems from MIT and is co-author of the book, What to Expect When You're Expecting Robots: The Future of Human-Robot Collaboration (Basic Books, 2020).ICAPS 2024 Keynote: Hector GeffnerICAPS2024-06-17 | ICAPS 2024 Keynote: Hector Geffner
Learning Representations to Act and Plan
Recent progress in deep learning and deep reinforcement learning (DRL) has been truly remarkable, yet two important problems remain: structural policy generalization and policy reuse. The first is about getting policies that generalize in a reliable way; the second is about getting policies that can be reused and combined in a flexible, goal-oriented manner. The two problems are studied in DRL but only experimentally, and the results are not clear and crisp. In our work, we have tackled these problems in a slightly different manner, developing languages for expressing general policies, and methods for learning them using combinatorial and DRL approaches. We have also developed languages for expressing and learning lifted action models, general subgoal structures (sketches), and hierarchical polices. In the talk, I'll present the main ideas and results, and open challenges. This is joint work with Blai Bonet, Simon Stahlberg, Dominik Drexler, and other members of the RLeap team at RWTH and LiU.
Short Bio
Hector Geffner is an Alexander von Humboldt Professor at RWTH Aachen University, Germany and a Guest Wallenberg Professor at Linköping University, Sweden. Before joining RWTH, he was an ICREA Research Professor at the Universitat Pompeu Fabra, Barcelona, Spain. Hector obtained a Ph.D. in Computer Science at UCLA in 1989 and then worked at the IBM T.J. Watson Research Center in New Work, and at the Universidad Simon Bolivar in Caracas. Distinctions for his work and the work of his team include the 1990 ACM Dissertation Award and three ICAPS Influential Paper Awards. Hector currently leads a project on representation learning for acting and planning (RLeap) funded by an ERC grant.ICAPS 2024: Tutorial 5ICAPS2024-06-11 | Orchestrating autonomous agents: Reinforcement Learning To Hierarchical Planning with COACH
Abstract: Live evaluation of cybersecurity defenses, or red team engagements, can be costly, difficult to commission, and inconsistent in scope, detail, and results. This high overhead prevents many organizations from fully using them despite their benefits. The objective of CALDERA is to enable automated assessment of a network's susceptibility to an adversary, essentially allowing an organization to see their network through the eyes of attackers on demand. CALDERA features an adversary model that maps to the MITRE ATT&CK® framework and an extensible planning system able to select and execute techniques. Inspired by automated planning methodologies, CALDERA provides a flexible, mature platform for developing adaptive and intelligent cyber agents.
Our tutorial encourages automated planning researchers to apply their skills to pressing cybersecurity issues. Towards this end, we introduce attendees to CALDERA, and use CALDERA to automate attacks against an enterprise network. Attendees learn how CALDERA gathers information and makes decisions, as well as how to modify those capabilities and run their own cybersecurity trials.ICAPS 2022: Tutorial on Planning & Scheduling and Quantum ComputingICAPS2022-07-12 | Presenters: Marco Baioletti, Angelo Oddi, Riccardo Rasconi
Abstract: Quantum Computing represents the next big step towards speed boost in computation, which promises major breakthroughs in several disciplines including Artificial Intelligence for the resolution of important classes of problems. Quantum algorithms process information stored in qubits, the basic memory unit of quantum processors, and quantum operations (called gates) are the building blocks of quantum algorithms. In order to be run on quantum computing hardware, quantum algorithms must be compiled into a set of elementary machine instructions (i.e., quantum gates). Since currently available quantum circuits suffer a number of technological problems such as noise and decoherence, it is important that the process that carries out the quantum computation be somehow adapted to the physical limitations of the quantum hardware of interest, by means of a proper compilation. This is the point in which AI-based P&S techniques can be of help to produce efficient compilation plans. On the other side, the speed-up promised by quantum technology may be greatly beneficial to solve P&S problems of more and more realistic size.ICAPS 2022: Tutorial on CraftBotsICAPS2022-07-12 | Presenters: Michael Cashmore, Liudvikas Nemiro
Abstract: The CraftBots simulation aims to be a new benchmark and competition environment for integrated planning and execution. Using a planning system effectively in the control of an agent acting in real-time poses a variety of challenges in integrating phases of planning with execution of plans. Integrated systems are developed with a focus on particular challenges, and it has been typically difficult to test, benchmark, and compare these systems. To do so requires a benchmark that has transparent and well-defined rules, and can be adapted to exhibit the problem characteristics of interest.
Craftbots is an accessible (requiring only python3 and numpy) and adaptive benchmark simulation for integrated planning and execution. The simulation can be configured to present a wide variety of different scenarios and exhibit different problem features, such as temporal uncertainty, non-deterministic action outcomes, partial observability, and limited communications.
In this tutorial we guide participants through hands-on coding exercises during which participants learn how to configure the scenario, implement a simple agent, and integrate the agent with the simulation API.ICAPS 2022: Tutorial on Representation Learning for Acting and PlanningICAPS2022-07-12 | Presenters: Blai Bonet, Hector Geffner
Abstract: In bottom-up approaches to representation learning, the learned representations are those that result from a deep neural net after training. In top-down approaches, representations are learned over a formal language with a known semantics, whether by deep learning or by any other method. There is a clean distinction between what representations need to be learned (e.g., in order to generalize), and how such representations are to be learned. The setting of action and planning provides a rich and challenging context for representation learning where the benefits of top-down approaches can be shown. Three central learning problems in planning are: learning representations of dynamics that generalize, learning policies that are general and apply to many instances, and learning the common subgoal structure of problems; what in reinforcement learning are called intrinsic rewards. In this tutorial, we look at languages developed to support these representations and methods developed for learning representations over such languages.ICAPS 2021 Dynamic Pickup and Delivery Problem from HuaweiICAPS2021-08-10 | Details: icaps21.icaps-conference.org/CompetitionsICAPS 2021 L2RPN Challenge with Trust: Learning to Run a Power NetworkICAPS2021-08-10 | Details: icaps21.icaps-conference.org/CompetitionsICAPS 2021 The Flatland Challenge: Multi-Agent Reinforcement Learning on TrainsICAPS2021-08-10 | Details: icaps21.icaps-conference.org/CompetitionsICAPS 2021 Competition on Automatic Reinforcement Learning for Dynamic JobShop Scheduling ProblemICAPS2021-08-10 | Details: icaps21.icaps-conference.org/CompetitionsICAPS 2020 Best Dissertation Award Talk - Kyle WrayICAPS2020-10-30 | Thesis Topic: Abstractions in Reasoning for Long-Term Autonomy
The dissertation is a well-rounded enterprise for designing and deploying control solutions for autonomous vehicles based on decision theoretic planning. It shows how to develop and successfully deploy such solutions, and presents novel contributions in different areas: hierarchical MDP/POMDP planning, design of controllers, multi-objective decision making, semi-autonomous systems, and scalable online decision making. For hierarchical planning, the thesis introduces policy networks that permit a unified and clear integration of subproblems and their solutions, resolving the problem of transferring control between subproblems. In design of controllers, the belief-infused finite-state controller for POMDPs are introduced. In multi-objective decision making problems, the thesis develops the topological MDP model together with scalable algorithms. In the semi-autonomous setting (SAS), the dissertation proposes a model for semi-autonomous that is used to decide when help from the human is required together with analyses of its properties in terms of safety and operation. Finally, in scalable online decision making, the dissertation presents a technique called MODIA that permits an effective solution for the so-called intersection problem for autonomous vehicles which presents itself when the vehicle arrives at an intersection where other uncontrollable entities are found.ICAPS 2020 - Community Socializing Entertainment ProgramICAPS2020-10-30 | Compilation of several "live" ICAPS music acts plus special interview, played at the ICAPS 2020 Community Socializing Event, by Lee McCluskey.ICAPS 2020 ThemeICAPS2020-10-30 | "Live" performance at the ICAPS 2020 Community Socializing Event by Lee McCluskey, Nick Hawes, Alan Lindsay, Mark Giuliano, and Jörg HoffmannICAPS 2020 - The Blocks of IpanemaICAPS2020-10-30 | "Live" performance at the ICAPS 2020 Community Socializing Event by Lee McCluskey, Mark Giuliano, and Jonah GiulianoICAPS 2020: DC Invited talk on Polanyi vs. PlanningICAPS2020-10-28 | Presenter: Subbarao Kambhampati (Arizona State University)
Invited Talk at the Doctoral Consortium at ICAPS 2020
Polanyi vs. Planning (Planning around AI's New Romance with Tacit Knowledge)ICAPS 2020: Invited talk on Getting the most out of your planner(s): from static to dynamic ...ICAPS2020-10-28 | Presenter: Frank Hutter (University of Freiburg)
Abstract: The many ingenious approaches underlying state-of-the-art planning systems tend to have complementary strengths; no single planner, heuristic, or parameter setting performs best in all situations, and machine learning can help improve performance. In the first part of this talk, I will survey proven meta-algorithmic approaches & success stories along these lines, including algorithm configuration, algorithm selection, algorithm schedules, and per-instance algorithm configuration. In the second part of the talk, I will then discuss a novel exciting meta-algorithmic framework dubbed dynamic algorithm configuration (DAC) that generalizes all of the meta-algorithmic approaches above and show a first case study of successfully applying DAC to improve planning performance. The DAC framework opens up many new opportunities, and I would be excited if the ICAPS community helps to explore them.ICAPS 2020: Tutorial on Certified Unsolvability in Classical Planning Part 2/2ICAPS2020-10-26 | Presenter: Salomé Eriksson (University of Basel) Co-Organizers: Gabriele Röger (University of Basel), Malte Helmert (University of Basel)
Abstract: Asserting correctness is an important step in both academic research and practical applications for increasing trust in planning systems. For example, it has been a long standing practice in the International Planning Competition to validate all plans with an independent validator. With the recent increased interest in unsolvable planning tasks however, new validation tools are needed that are capable of verifying unsolvability claims. This tutorial presents one such tool based on a proof system approach, where unsolvability of a specific task is proven by iteratively expanding a knowledge base and applying deduction rules. In addition to presenting the overall philosophy and architecture of the proof system, we focus on the question how planning systems can emit proofs verifiable by the proof system. Participants are expected to have basic knowledge about classical planning but do not need any other background.ICAPS 2020: Tutorial on Certified Unsolvability in Classical Planning Part 1/2ICAPS2020-10-26 | Presenter: Salomé Eriksson (University of Basel) Co-Organizers: Gabriele Röger (University of Basel), Malte Helmert (University of Basel)
Abstract: Asserting correctness is an important step in both academic research and practical applications for increasing trust in planning systems. For example, it has been a long standing practice in the International Planning Competition to validate all plans with an independent validator. With the recent increased interest in unsolvable planning tasks however, new validation tools are needed that are capable of verifying unsolvability claims. This tutorial presents one such tool based on a proof system approach, where unsolvability of a specific task is proven by iteratively expanding a knowledge base and applying deduction rules. In addition to presenting the overall philosophy and architecture of the proof system, we focus on the question how planning systems can emit proofs verifiable by the proof system. Participants are expected to have basic knowledge about classical planning but do not need any other background.ICAPS 2020: Tutorial on Evaluating Planners with Downward LabICAPS2020-10-23 | Presenter: Jendrik Seipp (University of Basel)
Abstract:
Downward Lab simplifies comparing planning algorithms experimentally on a single machine or on a computer cluster. Its primary use case is running experiments for planners built on top of Fast Downward, but it is also used for other classical and probabilistic planners. In this tutorial, I will explain how to set up experiments and make reports, talk about best practices for experimentation with Downward Lab and present solutions to common issues. At the end of the tutorial, we will have a Q&A session. The tutorial assumes basic familiarity with Python.ICAPS 2020: Tutorial on Epistemic PlanningICAPS2020-10-22 | Presenters: Thomas Bolander (Technical University of Denmark), Thorsten Engesser (University of Freiburg), Robert Mattmüller (University of Freiburg), Sheila McIlraith (University of Toronto)
Automated planning is of central concern in high-level symbolic AI research, with applications in logistics, robotics and service composition. Epistemic planning is the enrichment of automated planning with epistemic notions, including knowledge and beliefs, which not only refer to incomplete knowledge, but also beliefs about this knowledge. In general, single-agent epistemic planning considers the following problem: given my current state of knowledge, and a desirable state of knowledge, how do I get from one to the other? In multi-agent epistemic planning, the current and desirable states of knowledge might also refer to the states of knowledge of other agents, including higher-order knowledge like ensuring that agent a doesn’t get to know that agent b knows p.Single-agent epistemic planning is of central importance in settings where agents need to be able to reason about their own lack of knowledge and e.g. make plans of how to achieve the required knowledge. Multi-agent epistemic planning is essential for coordination and collaboration in multi-agent systems, where success can only be expected if agents are able to reason about the knowledge, uncertainty and capabilities of other agents. It is a relatively recent area of research, and is inherently multi-disciplinary involving the areas of automated planning, epistemic logic, and knowledge representation & reasoning. In order to achieve formalisms and systems for epistemic planning that are both expressive and practically efficient, it is necessary to combine state of the art from all three areas.ICAPS 2020: Tutorial on Causality, Creativity and Imagination: New Frontiers in PlanningICAPS2020-10-21 | Presenter: Sridhar Mahadevan (Adobe Research & University of Massachusetts Amherst)
The history of AI is inextricably linked with advances in the theory of planning. Beginning with the earliest work in AI on logical formulations of planning, such as STRIPS, to the more recent work on stochastic planning under uncertainty and reinforcement learning, formulations of planning have progressively become more sophisticated to meet the demands of real-world applications. Recent advances in machine learning, particularly deep reinforcement learning, have once again cast planning in a new light, enabling the development of agents that can plan in complex video games without the need for a priori models. In this tutorial, we explore new frontiers of planning that may emerge as a result of advances in other areas of AI, particularly deep learning models of imagination, such as generative adversarial networks, causal graphical models, and intrinsic motivation formulations of creativity. Humans exhibit a strong predisposition to imagine — to mentally transcend time, place, and circumstance — from an early age, an ability that is at the heart of all creative human activities, from art, literature, poetry, science, and technology. The ability to imagine is strongly connected to planning, as it is related to prediction of future states, and yet, imagination in humans is considerably more sophisticated than existing formulations of planning, such as those based on Markov decision processes. Imagination involves the construction of counterfactuals (e.g., what if Hilary Clinton had been elected President of the United States), as well as the elucidation of explanations (e.g., what is causing climate change?). At the core of human intelligence lies the notion of creativity, a capability that is prized among artists, scientists, and technological entrepreneurs. What is the relationship between creativity and planning? The tutorial will outline the key challenges in developing new formulations of planning that introduce causality, creativity, and imagination into the objectives of a deliberative agent. The tutorial will also elaborate novel connections between imagination-based planning and ongoing research in various areas, such as causality, deep learning, and transfer learning, and show why reformulating planning to include these additional capacities may transform AI in the coming decades.ICAPS 2020: Giacomo et al. on Imitation Learning over ...ICAPS2020-10-21 | ICAPS 2020 talk on the paper
Giuseppe De Giacomo, Marco Favorito, Luca Iocchi, Fabio Patrizi. Imitation Learning over Heterogeneous Agents with Restraining Bolts.
Proceedings of the Thirtieth International Conference on Automated Planning and Scheduling 2020.
(online available at aaai.org/Library/ICAPS/icaps20contents.php)ICAPS 2020: Industry session Planning And Scheduling For On-Board Aerospace ApplicationsICAPS2020-10-20 | Presenters: Edith Henry [SAFRAN] - Christophe Guettier [SAFRAN]
Abstract: Planning and Scheduling (P&S) find many applications in space and aeronautic domains, where autonomous solutions are growing quickly, as in ground transportation or in defense areas. Ranging from piloting assistance to autonomous unmanned systems, P&S has an in-depth impact on the system architecture which must face several challenges. The operational environment provides multiple sources of contingencies and uncertainties, due to the kinematic of surrounding objects, the lack of efficient situational awareness, the interactions with human beings and the potential hazards. In the assistance / autonomous system state of the art, the execution control must constantly adapt the plan to the reality, dealing with contingencies and uncertainty. In addition the lack of homogeneous dataset prevents from using approaches solely based on machine learning approach. However, aerospace systems must be fault-tolerant, and their design have to demonstrate correct execution with respect to various types of failures, placing in it a “safe” state (so called fail-safe mode), or able to continue the mission (so called fail-operational mode). In some cases it is possible to optimize fault recovery, such that the system becomes resilient. As a result of design process, some system constraints have to be applied to the embedded processing architecture, which must provide integrity and safety properties. The presentation reviews the impacts of fail-safe and fail-operational aerospace systems on planning models, solving algorithms and their integration environment.ICAPS 2020: Industry session Generic Optimization Engine For Solving Industrial Scheduling ProblemsICAPS2020-10-20 | Presenter: Philippe Laborie [IBM Data & AI]
Abstract: Classical scheduling problems (like job-shop or RCPSP) are among the most difficult problems studied in combinatorial optimization. Still, they are far from accounting for all the complexity of industrial scheduling applications. Since more than 20 years, our team at ILOG (now IBM) develops and integrates a large panel of techniques from AI (constraint programming, temporal reasoning, learning, ...) and OR (mathematical programming, graph algorithms, local search, ...) to solve our customers most complex scheduling problems. These works have lead to the design of CP Optimizer, a generic solver based on a very expressive (but still, quite concise) mathematical modeling language to formulate complex scheduling problems. The models are solved with an automatic search algorithm that is exact, efficient, robust and continuously improving. This talk gives a short overview of CP Optimizer.ICAPS 2020: Industry session Classical Planning To The RescueICAPS2020-10-20 | Presenter: Michael Katz [IBM Research AI]
Abstract: Planning, the task of finding a procedural course of action for declaratively described systems, is one of the oldest and well studied tasks in the field of Artificial Intelligence. Over the years, planning techniques have been applied to many real life applications. A few examples include but not limited to robotics, manufacturing, cyber security, diagnosis and remediation, logistics, transportation, and decision making in space. These applications often go beyond the classical setting in planning, requesting numeric features, state constraints, soft goals, non-deterministic effects, hierarchical compositional structure, etc. Further, even in cases that are close to classical planning, there are often no symbolic models available and, as a result, such models need to be constructed automatically from the existing data. Both these aspects pose a significant knowledge engineering and extraction challenge. Some of it, however, can be compiled away, some can be mitigated. In this talk, I will show how do we tackle problems with seemingly beyond classical features, using classical tools. Such tools include planning task transformations and reformulations, as well as top-k/top-quality/diverse planners. Using classical planners allows us to easily benefit from the progress of the planning community, constantly producing better classical tools, and solve problems of a large size.ICAPS 2020: Industry session Planning And Routing Tool For Smart LogisticsICAPS2020-10-20 | Presenters: Nastaran Rahmani [AntsRoute] - Ammar Oulamara [University of Lorraine]
Abstract: At AntsRoute, we are providing an optimization software that allows to solve different optimization problems in logistics. Our software not only facilitates decision making for daily, weekly, and/or monthly planning of ordinary user or a professional customer in an optimal manner but also it visualizes the obtained plans via a very user friendly graphical interface. The optimization library is developed in a generic way combining algorithms of operation research and artificial intelligence that can address different kind of Vehicle Routing Problems (VRP): Classic VRP, multi-depot VRP, Pickup and Delivery Problem (PDP). For each type of problems, any set of constraints for time window, capacity, vehicle duration, skills, and driver availabilities can be activated. The software can also help users to decide which date and time window is the optimal choice among the existing availabilities upon the arrival of a new order. Our so far experience shows that clients are looking for a software that provides a solution with high quality in a short amount of time. However, existing optimization algorithms cannot necessarily answer to all customer demands in a short amount of time. Besides, they are looking for a software that provides flexibility in their plannings : they prefer make changes over provided plans in real time.ICAPS 2020: Industry session Planning And Scheduling In Aerospace Applications With Simulators OnlyICAPS2020-10-20 | Presenter: Florent Teichteil-Königsbuch [Airbus]
Abstract: In many planning and scheduling aerospace applications that are investigated in the Airbus AI research group, there is no access to a white box model of the transition logics between states. For instance, planning aircraft paths in airways requires to evaluate the fuel burn and flight time between two successive waypoints, which can be only evaluated on-the-fly by an aircraft performance simulator as a function of current aircraft mass, speed, altitude and atmospheric conditions. At a lower level, planning continuous trajectories between successive waypoints for a real aircraft calls for a simulation of the dynamics of the aircraft based on a complex interaction between many avionic systems. For satellite mission planning applications, the dynamics of satellites and of the observed phenomena are usually so complex that they can only be accessed via simulators using current satellite's orientation and orbit position, weather conditions and predictions, and observation requests' status. Even worse, those simulators are usually very time-consuming, sometimes requiring several seconds of computation for a single simulation step, i.e. for generating a single transition of the planning problem. In addition, the state and action spaces are so huge that they cannot be enumerated beforehand, so as the transitions of the planning system. Those challenges impose to rethink the assumptions behind many classical algorithms from the community. We argue that the complexity of our aerospace applications resides more in the lack of transparent transition models and high combinatorics rather than on the expressivity of the underlying problem class which is generally quite simple. We will show how our Airbus AI research team, in collaboration with academic planning research groups, adapted some path planning, reinforcement learning, width-based planning and meta-heuristic algorithms to the challenges of real aerospace sequential decision-making with simulators only.ICAPS 2020: Tutorial on Regularization in Reinforcement LearningICAPS2020-10-20 | Presenter: Olivier Pietquin (Google Brain)
Abstract: Deep Reinforcement Learning (DRL) has recently experienced increasing interest after its success at playing video games such as Atari, DotA or Starcraft II as well as defeating grand masters at Go and Chess. However, many tasks remain hard to solve with DRL, even given almost unlimited compute power and simulation time. These tasks often share the common problem of being “hard exploration tasks”. In this tutorial, we will show how using demonstrations (even sub-optimal) can help in learning policies through different mechanisms such as imitation learning, inverse reinforcement learning, credit assignment and others.ICAPS 2020: Invited talk on Doing For Our Robots What Nature Did For UsICAPS2020-10-20 | Presenter: Leslie Pack Kaelbling (Massachusetts Institute of Technology)
Abstract: We, as robot engineers, have to think hard about our role in the design of robots and how it interacts with learning, both in "the factory" (that is, at engineering time) and in "the wild" (that is, when the robot is delivered to a customer). I will share some general thoughts about the strategies for robot design and then talk in detail about some work I have been involved in, both in the design of an overall architecture for an intelligent robot and in strategies for learning to integrate new skills into the repertoire of an already competent robot.ICAPS 2020: Invited Talk on Learning Planning Representations From Traces Via SATICAPS2020-10-20 | Presenter: Blai Bonet (Universidad Simón Bolívar)
Abstract: Recent work in planning, generalized planning and representation learning is concerned with the problem of learning symbolic representations from traces over small instances. Such representations can be used for different purposes such as to find plans for bigger unseen instances, to do plan and goal recognition, as building blocks for constructing more complex models, etc. In this talk, I will show how SAT solvers can be used to learn first-order STRIPS representations from purely non-symbolic traces, and to learn representations for generalized planning, based on qualitative numerical planning, from symbolic and non-symbolic traces. A common denominator when learning representations from non-symbolic traces is a crisp graph-theoretical problem that is cast as a search problem in a combinatorial space and solved with the help of SAT solvers.ICAPS 2020: Barták et al. on Multi-agent path finding on real robotsICAPS2020-10-19 | ICAPS 2020 talk on the paper
Roman Barták, Jiří Švancara, Věra Škopková, David Nohejl, Ivan Krasičenko. Multi-agent path finding on real robots.
Proceedings of the Thirtieth International Conference on Automated Planning and Scheduling 2020.
Mohan Sridharan, Michael Gelfond, Jeremy Wyatt, Shiqi Zhang. REBA: A Refinement-Based Architecture for Knowledge Representation and Reasoning in Robotics.
Proceedings of the Thirtieth International Conference on Automated Planning and Scheduling 2020.
Aye Phyu Phyu Aung, Xinrun Wang, Bo An, Xiaoli Li. We Mind Your Well-Being: Preventing Depression in Uncertain Social Networks by Sequential Interventions.
Proceedings of the Thirtieth International Conference on Automated Planning and Scheduling 2020.