Kenote@IJCCI 2017: Towards Lifelong Learning in Optimisation Algorithms

Start date and time

Thursday 2 November 2017

Location

Funchal, Madeira

Optimisation is an important activity for many businesses, providing better, faster, cheaper solutions to problems in areas including scheduling of people and processes, routing of vehicles and packing of containers. Metaheuristic algorithms provide a pragmatic way to tackle optimisation, providing high-quality solutions in reasonable time. Unfortunately, selection and tuning of an appropriate algorithm can difficult, often requiring an expert to design the algorithm, a software engineer to implement it, and finally application of automated tuning processes to refine the chosen algorithm. This is not only costly, requiring significant human-effort, but also results in software which can quickly become obsolete when it no longer matches the goals of a company or if the characteristics of the optimisation problems being solved changed substantially. Unlike human-beings, optimisation software is currently unable to adapt to changing scenarios or autonomously improve its behaviour over time as it learns from experience.
To counter this, I will propose the life-long learning optimisation system (L2O) which when faced with a continual stream of problems to optimise, refines an existing set of algorithms so that they improve over time as they are exposed to more examples, and automatically generates new algorithms when faced with problem instances that are completely different from those seen before. The approach is inspired by ideas from the operation of the natural immune system, which exhibits many properties of a life-long learning system that can be exploited computationally, and uses genetic programming to automatically generate new algorithms. I will give a brief overview of the immune system, focusing on highlighting its relevant computational properties and then show how it can be used to construct a lifelong learning optimisation system. The system is shown to adapt to new problems, exhibit memory, and produce efficient and effective solutions when tested in both the bin-packing and scheduling domains, representing a paradigm shift in the way we think about optimisation.