Representation learning for acting and planning: A top-down approach

An ICAPS'22 Tutorial

(quarter day)

June 14, 2022


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 reinforcemenet learning are called intrinsic rewards. In the tutorial, we will look at languages developed to support these representations and methods developed for learning representations over such languages.

Target audience

Students and researchers interested in representation learning, in the setting of actions and planning. Basic knowledge of AI and planning is assumed.


  • PART 0. Introduction. General idea: Learning representations over languages
    • Bottom up vs. top down approaches to representation learning
    • Goals and focus of the tutorial
  • Part I. Representations: Dynamics, policies, subgoals
    • Representing Dynamics: Syntax and Semantics (Lifted STRIPS, PDDL, SAS+, etc)
    • Representing General Policies: Feature-based representations. General policies and QNPs. General policies and width. Transfer learning and one-shot generalization.
    • Representing Subgoal Structure: Macro tables, HTNs, Sketches, Problem and sketch width.
  • Part II: Learning representations of dynamics, policies, subgoals
    • Learning general action dynamics given: a) the domain predicates and objects, b) sequences of ground actions, c) action labels, d) states as images or parsed images.
    • Learning general policies: a) Find most compact general policies representation that solves sample problems, b) Learning feature-based policies and feature-based value functions, c) Synthesis of general policies from specifications vs. learning them from examples.
    • Learning subgoal structures and hierarchies: Learning HTNs. Learning sketches.
  • Part III: Learning general dynamics and policies with deep (reinforcement) learning
    • Formulations, Results, Open Problems, Challenges
    • Features, Finite-Variable Logic, and GNNs
  • Part IV: Future, Challenges
    • Continuous domains, spaces, actions, and time
    • Stochastic and non-deterministic domains
    • Extended temporal Goals


Blai Bonet

Blai Bonet: Blai is a professor in the computer science department at Universidad Simon Bolivar, Venezuela. He received his Ph.D. degree in computer science in 2004 from the University of California, Los Angeles. His research interests are in the areas of automated planning, heuristic search and knowledge representation. He has received several best paper awards or honorable mentions, including the 2009 and 2014 ICAPS Influential Paper Awards, and he is a co-author of the book ”A Concise Introduction to Models and Methods for Automated Planning”. Dr. Bonet has served as associate editor of Artificial Intelligence and the Journal of Artificial Intelligence Research, conference co-chair of ICAPS-12, program co-chair of AAAI-15, and has been a member of the Executive Council for ICAPS and AAAI.

Hector Geffner

Hector Geffner: Hector got his Ph.D at UCLA with a dissertation that was co-winner of the 1990 ACM Dissertation Award. He then worked as Staff Research Member at the IBM T.J. Watson Research Center in NY, USA and at the Universidad Simon Bolivar, in Caracas, Venezuela. Since 2001, he is a researcher at ICREA and a professor at the Universitat Pompeu Fabra, Barcelona, and since 2019, a Wallenberg Guest Professor at Linköping University, Sweden. Hector is a Fellow of AAAI and EurAI, and the author of the book “Default Reasoning: Causal and Conditional Theories” , MIT Press, 1992, editor of ”Heuristics, Probability, and Causality: a Tribute to Judea Pearl” along with R. Dechter and Joe Halpern, College Publications, 2010, and author of ”A Concise Introduction to Models and Methods for Automated Planning” with Blai Bonet, Morgan and Claypool, 2013. He leads a project on representation learning for planning (RLeap), funded by an Advanced ERC grant (2020–2025).