ORAHS2024
Abstract Submission

190. Charting the Pareto frontier: Multi-objective optimization for platelet inventory management using neuroevolution

Contributed abstract in session FB-4: Artificial Intelligence, stream Regular talks.

Friday, 11:00-12:30
Room: Room S3

Authors (first author is the speaker)

1. Joseph Farrington
Institute of Health Informatics, University College London
2. Kezhi Li
Institue of Health Informatics, UCL
3. Martin Utley
Clinical Operational Research Unit, University College London

Abstract

Managing perishable inventory, such as platelets, requires balancing multiple objectives. In the blood product management literature, notional costs are often used to weight the relative importance of avoiding shortages and minimizing wastage because the consequences are not just monetary. The notional costs do not have an empirical basis and do not necessarily accurately reflect the preferences of decision makers. Estimating the Pareto frontier of feasible policies can illustrate potential trade-offs between objectives and support discussions with decision-makers.

We used evolutionary algorithms to fit the parameters of neural networks representing both replenishment and issuing policies. This enabled us to learn complex policies and use key performance indicators, including wastage and service level, as the objectives in multi-objective optimization algorithms to estimate the Pareto frontier. We compared the performance of these policies with heuristic replenishment and issuing policies and with an estimate of the optimal policies computed over a range of notional cost ratios using dynamic programming. Our findings show that the neuroevolution approach is competitive with other approaches and outperforms them in a number of instances.

In this presentation, we will also explain how this work was facilitated by using the Python library JAX to run algorithms in parallel on affordable, consumer-grade GPU hardware.

Keywords

Status: accepted

Back to the list of papers