AAAI 2022 Tutorial on AI Planning: Theory and Practice
Brief description
The tutorial provides a theoretical background on AI Planning and introduces some of the existing tools, as well as
existing applications that were tackled with these tools. We show how to use the tools on an example application.
The participants will get the knowledge and hands-on experience necessary to start using AI Planning tools in their
applications.
Overview
AI Planning is a long standing sub-area of AI, dealing with sequential
decision making, a sister field to Reinforcement Learning. Mature industrial
applications of planning technology can be seen in various fields, such as
dialog systems, cybersecurity, transportation and logistics, as well as IT.
While model-based planning tools allow to solve problems of practical size, applying AI
Planning research in practice faces several challenges, preventing its
widespread use. The alternative of using model-free methods, however, often
proves infeasible for real life size problems.
The aim of this tutorial is to provide the audience with the
necessary theoretical background knowledge, as well as hands-on experience to
allow for using planning tools for solving everyday challenges.
In this tutorial, we will provide an overview of the field of planning, including recent advances in the field. We
will then dive into three challenges: (1) modeling — how to represent, extract, and learn the knowledge; (2) theory and tools — formal definition of the computational problems and how to solve these problems; (3) practice — using AI Planning in end-to-end applications. We will
have a hands-on session to exemplify the use of planning tools for solving an example application. Our aim is to
give AAAI attendees the necessary means for using AI planning tools in their applications.
Target Audience
Potential target audience are researchers in AI that want to know more
about planning, typical AAAI attendees. No prior knowledge is necessary,
since we plan (pun intended) to cover all the necessary background. In
particular, we hope that researchers with no prior exposure to AI Planning
will find our tutorial interesting and motivating for their work.
Why Should I Attend
Our main goal is to give the researchers in AI that are interested in
sequential decision making a basic knowledge about the field of AI planning.
We will provide you with the tools to solve various real life size sequential decision making problems.
We will cover several applications of AI Planning, challenges, and
solutions and give the attendees a hands-on experience of using planning tools in one of these applications.
- Introduction
[slides]
- What is AI Planning
- Why? Motivation of this tutorial — applications
- Why AI Planning is important
- How to spot a planning problem — why these applications are planning problems
- AI Planning, how it works
- Plan for this tutorial (theory, modeling, practice)
- Theory and planning tools
[slides]
- What is AI Planning: models and languages
- Models and languages
- Planning and RL
- Computational problems
- Solving classical planning
- History
- Planning as heuristic search
- Heuristics for classical planning
- Search pruning techniques
- Planners and planning competitions
- International planning competition
- Major planning toolkits/systems/families
- Non-IPC planners and tools
- Modeling
[slides]
- Modeling challenges
- Relationship to planning
- Overview of our solutions
- Automating machine learning pipeline generation
- Hypothesis generation problem
- Scenario planning for enterprise risk management
- Planning for dialog
- Overview of other approaches
- Modeling summary
- Practice: Tools and applications
[slides]
- Challenges in using AI Planning in applications
- Pre/post processing
- Hands on: automating ML pipeline generation/exploration (instructions here and more details)
- Problem definition
- Modeling
- Transformation (HTN to classical planning transformation)
- Additional constraints
- Computation of plans (top-k planning)
- Postprocessing of the plans to pipelines
- Pipeline accuracy and its translation to action costs
- Quick overview of other applications:
- Automated analytic composition
- Data science automation
- Hypothesis generation
- Scenario planning
- Summary and Q&A
[slides]
- Lessons learned
- Improvements/future work
- Instructions for the hands-on part and the code/notebooks you need are on Github
- Some planning tools we will be using in this tutorial:
Useful links
Major Planning Toolkits/Systems/Families
Fast-Forward (better known as FF)
Fast Downward
Lightweight Automated Planning ToolKiT (LAPKT)
LPG
SHOP2
OPTIC.
Non-IPC Planners and Tools
Planning service for cost-optimal, agile, satisficing, top-k,
top-quality, diverse planning, also available as
docker image
Planner in the cloud and a collection of
tools
Forbid-Iterative Collection of planners for top-k, top-quality,
diverse planning
Top-k planners K* and
SymK
OSP planners A*-based and
symbolic- search-based
FOND planner PRP
Pyperplan Lightweight python-based planner developed for educational
purposes
Other Major Community Efforts
Slack
workspace
ICAPS web site
Community GitHub
PlanUtils
Planning Wiki (initial effort), including a list of planners.
Previous Tutorials