Keep Learning: Towards Optimisers That Continually Improve and/or Adapt
Overview
Combinatorial problems are ubiquitous across many sectors, delivering optimised solutions can lead to considerable economic benefits in many fields. In a typical scenario, instances arrive in a continual stream and a solution needs to be quickly produced. Although there are many well-known approaches to developing optimisation algorithms, most suffer from a problem that is now becoming apparent across the breadth of Artificial Intelligence: systems are limited to performing well on data that is similar to that encountered in their design process, and are unable to adapt when encountering situations outside of their original programming.
For real-world optimisation this is particularly problematic. If optimisers are trained in a one-off process then deployed, the system remains static --- despite the fact that optimisation occurs in a dynamic world of changing instance characteristics, changing user-requirements and changes in operating environments that influence solution quality (e.g. changes in staff availability, breakdowns in a factory, or traffic in a city). Such changes may be either gradual, or sudden. In the best case this leads to systems that deliver sub-optimal performance, while at worst, systems that are completely unfit for purpose. Moreover, a system that does not adapt wastes an obvious opportunity to improve its own performance over time as it solves more and more instances.
The goal of this workshop is to discuss mechanisms by which optimisers can “keep on learning”. This includes mechanisms to enable an optimisation system to:
Improve with practice as it solves more and more instances
Learn & adapt from its experience of solving problem instances
Detect drift in instance characteristics and respond accordingly, e.g. by tuning solvers and/or models
Detect “surprise” in instance characteristics and respond accordingly, e.g. generation of new solvers
Predict empty regions of an instance-space where future instances might appear; generate new synthetic instances in this space to provide training data for solvers
Learn across multiple domains, e.g. transfer learning
Learn to optimise in unseen domains
Developing such a system will likely require an interdisciplinary approach that mixes machine-learning and optimisation techniques. The workshop solicits short papers that address mechanisms by which any of the above can be achieved. We also invite short position papers that do not contain results but propose novel avenues of work might enable the creation of life-long learners.
Topics of Interest
Possible topics include but are not limited to:
Per-instance Algorithm Selection
Developing dynamic algorithm portfolios
Algorithm Generation
Algorithm Tuning
Cross-Domain and/or Multi-Task Optimisation
Methods for Warm-Starting Optimisers
Methods for detecting change in instance characteristics
Feature-generation and selection
Synthetic Instance generation
Creating instance space maps
Keynote Speakers
TBA