Instance Specific Algorithm Selection and Portfolios

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2646 papers861 conceptsfrozen 2026-10-04

In computational science, the "No Free Lunch" theorem is a practical reality: no single algorithm can achieve peak performance across every possible instance of a problem. Whether solving Boolean Satisfiability (SAT), training deep neural networks, or calculating fluid dynamics, researchers consistently find that an algorithm that dominates on one dataset may fail catastrophically on another.

The research direction of Algorithm Selection and Portfolios moves away from the search for a "universal" solver. Instead, it focuses on building meta-systems that analyze the specific characteristics of a problem instance and automatically select—or even synthesize—the most efficient tool for that specific job.

The Portfolio Advantage

The core intuition behind this field is that a diverse collection of specialized "weak" solvers can outperform a single "strong" generalist. If we have a portfolio of algorithms with complementary strengths, the challenge shifts from algorithm design to a selection problem.

Early frameworks like Hydra, developed by Lin Xu, Holger H. Hoos, and Kevin Leyton-Brown, demonstrated that we don't even need human-designed diversity. Hydra automatically configures new solvers to specifically target the "weak spots" of an existing portfolio, iteratively filling the gaps in performance across a heterogeneous problem space.

This approach is particularly effective in combinatorial optimization. As Marius Lindauer and colleagues showed with AutoFolio, the selection strategy itself must be flexible. Different domains might require simple regression to predict runtimes, while others benefit from complex pairwise voting between candidate solvers.

Mapping Problem Features to Solver Strengths

To pick the right tool, a system must first "sense" the problem. This is done by extracting features—numerical descriptors of the problem's structure.

  • Numerical Computing: In solving Partial Differential Equations, Anupam Joshi and John R. Rice used neuro-fuzzy systems to detect mathematical properties like oscillations or singularities, which dictate whether a specific numerical solver will be stable or efficient.
  • Ontology Alignment: The RiMOM system, described by Juanzi Li and Jie Tang, estimates the "vibe" of data—linguistic versus structural—before deciding which alignment strategy to deploy.
  • Scientific ML: Recent work on SciML Agents by Xiang Zheng uses Large Language Models to identify properties like "stiffness" in differential equations, ensuring the agent writes code for an implicit solver when stability is at risk.

By treating the problem instance as a point in a high-dimensional feature space, these systems learn a mapping: if the problem looks like X, use algorithm Y.

Meta Learning and Optimizer Amalgamation

The proliferation of deep learning optimizers (Adam, Lion, Adan, SGD) has created a "crowded valley" where practitioners often default to Adam out of habit. However, researchers like Tianshu Huang and Tianlong Chen suggest we shouldn't have to choose.

Their work on Optimizer Amalgamation distills multiple "teacher" optimizers into a single "student" that adaptively switches strategies per-parameter and per-iteration. Similarly, Kaan Ozkara and colleagues proposed MADA, which uses hyper-gradient descent to navigate the "optimizer zoo" automatically.

This shift moves the burden of expertise from the human to the meta-algorithm. Instead of a researcher spending weeks tuning a single solver, a meta-learning system like TUPSO, proposed by Eitan Menahem and Lior Rokach, can combine multiple one-class classifiers to ensure robust performance even when traditional evaluation metrics are unavailable.

Why It Works: Exploiting Complementarity

The success of these methods relies on complementarity. In many NP-hard domains, the "hard" instances for Algorithm A are often "easy" for Algorithm B. Chengrui Gao and Ke Xue demonstrated this in combinatorial optimization, where different neural solvers excel on different problem geometries.

By using a meta-selection approach, we essentially "short-circuit" the worst-case complexity of any single algorithm. We are no longer limited by the performance of our best general-purpose tool, but rather by our ability to correctly identify which specialist is best suited for the current task. As benchmarks like the Open Source AutoML Benchmark by Pieter Gijsbers and Erin LeDell show, the "best" tool is a moving target—and automated selection is the only way to hit it consistently.

Papers in this version

Showing 60 of 646 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 457 papers
  2. Version 1 · 2026-10-04 · 646 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_algorithm_selection_problem_a2c10564,
  title = {algorithm selection problem},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/algorithm-selection-problem/rdr_a2c10564c71847768d58f22477bd4ddf},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_instance_specific_algorithm_selection_an_a2c10564_v1,
  title = {Instance Specific Algorithm Selection and Portfolios},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/instance-specific-algorithm-selection-and-portfolios/rdr_a2c10564c71847768d58f22477bd4ddf/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}