ORCID

Document Type

Conference Proceeding

Abstract

The Cardinality Constrained Portfolio Optimisation Problem (CCPOP) is an NP-hard challenge in finance, requiring simultaneous minimisation of risk and maximisation of return under strict cardinality and weight constraints. Traditional Multi-Objective Evolutionary Algorithms (MOEAs) approximate the entire Pareto front and often struggle to escape local optima in the search space. This paper proposes a framework, Hybrid Adaptive Preference Reinforcement Learning guided by Iterated Local Search, integrated into NSGA-II and driven by synthetic investor profiles via an RL-inspired adaptive selection mechanism. In this mechanism, an embedded agent learns to switch dynamically between diverse user profiles (conservative, neutral, aggressive, risk taker) based on immediate local search improvements. By modelling the local search as a sequential decision process, the agent employs a weighted average estimator to track the value of each preference profile and optimise the search trajectory. Applied to the CCPOP on benchmark datasets, our approach demonstrates that the adaptive mechanism significantly improves convergence and diversity compared to static preference models, offering a robust framework for preference-guided portfolio search.

Publication Date

2026-08-23

Event

19th International Conference on Parallel Problem Solving from Nature, PPSN 2026

Publication Title

Parallel Problem Solving from Nature – PPSN 19 - 19th International Conference, PPSN 2026, Proceedings

Publisher

Springer Science and Business Media Deutschland GmbH

ISBN

9783032362223

ISSN

0302-9743

Deposit Date

2026-11-16

Keywords

Combinatorial Problems, Evolutionary Algorithms, Machine Learning, Multi-objective Optimisation, Reinforcement Learning

Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

First Page

183

Last Page

198

Share

COinS