[Paper Review] Multi-Information Source Optimization with General Model Discrepancies

Matthias Poloczek, Jialei Wang+1 more|arXiv (Cornell University)|Mar 1, 2016

Advanced Multi-Objective Optimization Algorithms22 references3 citations

TL;DR

This paper proposes a novel optimization algorithm for multi-information source problems with model discrepancies, using a rigorous value-of-information analysis to balance exploration cost and expected benefit. It outperforms state-of-the-art methods by achieving higher objective values with reduced exploration cost.

ABSTRACT

In the multi-information source optimization problem our goal is to optimize a complex design. However, we only have indirect access to the objective value of any design via information sources that are subject to model discrepancy, i.e. whose internal model inherently deviates from reality. We present a novel algorithm that is based on a rigorous mathematical treatment of the uncertainties arising from the model discrepancies. Its optimization decisions rely on a stringent value of information analysis that trades off the predicted benefit and its cost. We conduct an experimental evaluation that demonstrates that the method consistently outperforms other state-of-the-art techniques: it finds designs of considerably higher objective value and additionally inflicts less cost in the exploration process.

Motivation & Objective

• To address the challenge of optimizing complex designs when objective values are only accessible through multiple information sources with inherent model discrepancies.
• To develop a decision-making framework that rigorously accounts for uncertainties arising from model discrepancies across sources.
• To minimize the cost of exploration while maximizing the expected improvement in objective value through informed source selection.

Proposed method

• The method employs a mathematical treatment of uncertainties due to model discrepancies across information sources.
• It uses a value-of-information analysis to quantify the expected benefit of querying each source relative to its cost.
• The algorithm dynamically selects information sources based on a trade-off between predicted improvement and acquisition cost.
• It models discrepancies as stochastic deviations from reality and integrates them into a probabilistic optimization framework.
• The approach combines Bayesian inference with cost-aware acquisition functions to prioritize high-value, low-cost queries.

Experimental results

Research questions

• RQ1 How can we effectively optimize complex designs when objective evaluations are only available through multiple, inaccurate information sources?
• RQ2 What is the optimal trade-off between exploration cost and expected improvement in objective value under model discrepancies?
• RQ3 How can uncertainty from model discrepancies be rigorously modeled and incorporated into the optimization decision process?

Key findings

• The proposed method consistently achieves higher objective values compared to state-of-the-art techniques in multi-information source optimization.
• It reduces the overall cost of the exploration process, requiring fewer and more strategically selected queries.
• The value-of-information analysis enables more efficient source selection, leading to faster convergence to high-quality designs.
• The method demonstrates robust performance across diverse problem instances with varying levels of model discrepancy.