acceptodds
Under review as a conference paper at ICLR 2027

CMDO: A Cognitive Memory-Driven Optimization Algorithm for Adaptive Population-Based Search

Abstract

Population-based optimization methods often use previous search information through successful solutions, parameter adaptation, or operator performance, but they rarely retain the context in which a search behavior succeeded or failed. We introduce Cognitive Memory-Driven Optimization (CMDO), a derivative-free population-based optimizer that represents experience as the relationship between search context, search behavior, and observed outcome. CMDO organizes these experiences across working, episodic, and consolidated memory, retrieves them according to similarity with the current search state, and uses both positive and negative evidence to guide subsequent search. Retrieved experience does not replay previous candidate locations; instead, it selects search recipes that are reconstructed from the current population through exploratory, directed, and local search behaviors with adaptive search geometry. We evaluate CMDO on selected Blackbox Optimization Benchmarking test suite on COCO (BBOB/COCO) and Congress on Evolutionary Computation 2017 (CEC2017) problems against DE, CMA-ES, SHADE, GWO, HHO, and ORCA, and further study its application to seven-parameter photovoltaic model estimation using measured current–voltage data. The results show problem-dependent but competitive optimization performance, including the lowest median error among the compared methods on CEC2017 F10. More importantly, analysis of the search traces shows that context-dependent recall changes the distribution of executed search behaviors, while unsuccessful experiences remain available as negative evidence for later decisions, showing that accumulated experience directly shapes subsequent search behavior. These results support the use of explicit context–behavior–outcome memory as an active mechanism for controlling population-based search.

Primary area:
optimization
Keywords:
Cognitive memory, black-box optimization, contextual retrieval, adaptive search, derivative-free optimization.

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_x47COpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Cognitive Memory-Driven Optimization (CMDO), a derivative-free, population-based optimizer that explicitly stores and reuses “experiences” in the form of (context, search recipe, reward). CMDO organizes these experiences across working, episodic, and consolidated memory, retrieves them by contextual similarity, and uses both positive and negative evidence to bias operator/parameter selection for future steps. The paper reports competitive results on a small subset of BBOB (D=5) and CEC2017 (D=10) functions and on a photovoltaic parameter-estimation task, and presents analyses suggesting that context-conditioned recall measurably reshapes the distribution of executed search behaviors.

Critical strengths
  1. The algorithmic components are well structured and explained, including context features, retrieval, negative evidence, and consolidation, with an appendix that provides more precise definitions and pseudocode.

  2. The memory mechanisms are described with enough intuition to understand how they influence search choices.

Critical weaknesses
  1. The context vector and similarity metric are hand-crafted and may be brittle; sensitivity to h, λ, ρ, top-k, and memory capacities is not explored.

  2. Only three search operators are considered; it remains unclear how the memory controller scales when the operator family is larger/more diverse.

  3. No complexity/overhead analysis is provided; the cost of retrieval/consolidation and its impact on wall-clock efficiency is not discussed.

  4. Baseline choices do not include state-of-the-art CMA/DE variants or portfolios (e.g., BIPOP-CMA-ES, LSHADE/NL-SHADE-RSP/jSO, NGOpt-wizard, Shiwa, MSC‑CMA‑ES). Reported SHADE/CMA-ES performance looks surprisingly weak in some cells, raising concerns about configuration fairness.

  5. The paper does not adequately position itself relative to hyper-heuristics and algorithm selection/portfolio controllers for black-box optimization (e.g., Shiwa/NGOpt; active/portfolio selection with warm-starts and bandits).

Questions for authors

see weakness

Reviewer_b2PvOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Cognitive Memory-Driven Optimization (CMDO), a derivative-free population-based optimization algorithm that uses historical search experiences to guide subsequent search decisions. The method represents each experience through three components: search context, search recipe, and observed reward. These experiences are maintained in a three-level memory architecture consisting of working, episodic, and consolidated memory. Based on contextual similarity, CMDO retrieves relevant experiences and uses both positive and negative outcomes to adaptively select among exploratory, directed, and local search operators. The proposed method is evaluated on selected BBOB and CEC2017 benchmark functions and a photovoltaic parameter-estimation problem. The results demonstrate competitive performance on some problems and show that memory retrieval influences the distribution of selected search operators. However, the effectiveness and distinct contributions of the proposed memory mechanisms remain insufficiently validated.

Critical strengths
  1. The paper investigates a relevant research problem in population-based optimization, namely how historical search experience can be represented and reused to improve subsequent search decisions. The context–behavior–outcome formulation provides an intuitive framework for incorporating contextual information into adaptive search.

  2. The proposed method integrates positive and negative search experiences into its memory retrieval and operator-selection mechanisms. This is a potentially useful design choice, as unsuccessful search experiences may provide information for avoiding ineffective search behaviors under similar conditions.

  3. The authors evaluate the proposed approach on both numerical optimization benchmarks and a photovoltaic parameter-estimation problem. The results suggest that the method can achieve competitive performance in certain cases, and the analysis of recalled search behaviors provides preliminary evidence that the memory mechanism influences search decisions.

Critical weaknesses
  1. The methodological novelty is not sufficiently established. The central idea of associating search states with operator performance is closely related to adaptive operator selection, reinforcement learning-based search control, and success-history-based parameter adaptation. Although CMDO introduces a context–behavior–outcome memory representation, the paper does not clearly explain what fundamental methodological advantage this representation provides over existing approaches. The authors should strengthen the comparison with relevant methods and clarify the distinctive contributions beyond organizing existing adaptation mechanisms into a hierarchical memory structure.

  2. The proposed memory architecture lacks sufficient justification. The working, episodic, and consolidated memory levels are motivated by cognitive memory concepts, but their computational roles and necessity are not convincingly established. In particular, it remains unclear why three memory levels are required, how their interactions improve optimization, and whether a simpler context-conditioned experience archive could achieve similar performance. The choices of memory capacities, consolidation criteria, and forgetting mechanisms also appear largely heuristic.

  3. The effectiveness of the core memory mechanism is not adequately validated. The experimental analysis mainly demonstrates that recalled recipes have different operator-selection distributions from freshly generated recipes. However, this observation alone does not establish that contextual retrieval improves optimization performance. More importantly, the paper lacks systematic ablation experiments comparing CMDO with variants without memory, without contextual retrieval, without negative experience, and without memory consolidation. These comparisons are essential for substantiating the main contributions.

  4. The experimental evaluation is too limited to support the claimed generality. Only three BBOB functions and three CEC2017 functions are evaluated, with dimensions of 5 and 10, respectively. Furthermore, the numerical results are reported using only three random seeds, which is insufficient for reliable statistical comparisons. A more comprehensive evaluation across benchmark functions, problem dimensions, and random seeds is necessary to demonstrate robustness and generalization.

  5. The optimization performance does not convincingly demonstrate an advantage over established methods. According to Table 1, CMDO achieves the best result only on CEC2017 F10 among the six reported numerical problems. On several other problems, established methods such as DE, CMA-ES, GWO, and ORCA perform better. Similarly, in the photovoltaic parameter-estimation experiment, DE substantially outperforms CMDO. While a new method does not need to outperform every baseline, the authors should provide stronger evidence identifying the problem characteristics under which the proposed memory mechanism offers meaningful advantages.

  6. The selection of baseline algorithms should be strengthened. Although the paper includes several established population-based optimizers, the comparisons do not sufficiently address the closest methodological alternatives, particularly adaptive operator selection and reinforcement learning-based evolutionary optimization. Since the main contribution concerns context-dependent selection and adaptation of search behaviors, comparisons with these approaches would be more informative than comparisons with several conventional metaheuristic algorithms.

  7. The computational complexity and scalability of the memory mechanism are insufficiently investigated. CMDO introduces additional operations for context construction, memory retrieval, similarity evaluation, negative-evidence aggregation, and memory consolidation. Although the memory capacities are bounded, the paper does not quantify the computational overhead or demonstrate how these mechanisms scale with problem dimensionality and population size. Runtime comparisons and scalability experiments would be valuable, particularly for expensive or large-scale optimization problems.

Questions for authors

please refer to the summary of weakness of this work.

Reviewer_38DpOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a new numerical optimisation algorithm called CMDO. The authors claim the design of the algorithm takes into account past search experience.

The algorithm is compared empirically with other algorithms on the BBOB and CEC2017 benchmarks. The paper also illustrates the algorithm on a photovoltaic parameter estimation problem.

Critical strengths

None

Critical weaknesses

The presentation is inadequate: The algorithm is not described properly with unambiguous pseudo-code.

The results lack significance: There is insufficient experimental evidence that CMDO is a well-performing algorithm. For several benchmark problems, the CMDO performed worse than the baseline algorithms.

The paper lacks novelty: The paper states the question: what should an optimizer remember? This is hardly a novel question. Any design of an optimisation algorithm must decide how information from evaluation of search points should be incorporated into the updated state of the algorithm.

The methodology is weak: There is insufficient statistical treatment of the experimental data. E.g., in Table 1, what variance was observed in the data?

Questions for authors

Given the numerous problems identified above, I do not have any questions that could materially affect my assessment.

1 comment

The author has not responded to my review as of Nov 17th. I therefore maintain my score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.