Source-linked AI summary
CMDO: A Cognitive Memory-Driven Optimization Algorithm for Adaptive Population-Based Search
Mohammed Yusuf Mujawar, Shahram Rahimi, Noorbakhsh Amiri Golilarz
TL;DR
Existing optimizers often use past solutions, parameter adaptation, or operator performance without retaining the context in which a behavior succeeded or failed. CMDO stores and retrieves context–behavior–outcome experience to shape new search actions. It shows competitive, problem-dependent performance, while trace analysis indicates recalled experience changes executed behavior distributions.
Problem
Existing population-based optimizers use past solutions, parameter adaptation, or operator performance, but rarely retain the context in which search behavior succeeded or failed.
Source: §1 · pp. 1–2Method
CMDO stores context–behavior–outcome experiences across working, episodic, and consolidated memory, retrieving them by contextual similarity to guide future search.
Source: §1 · p. 2Results
Across selected BBOB and CEC2017 problems and photovoltaic estimation, CMDO shows competitive, problem-dependent performance, and recalled experience changes the distribution of executed search behaviors.
Source: §4.5 · p. 9, §5 · p. 10Takeaways & Limitations
The results support explicit context–behavior–outcome memory as an active mechanism for controlling population-based search.
Source: §5 · p. 10The stated optimization setting is bounded continuous optimization.
Source: §3.1 · p. 4
Abstract
from arXiv · showhide
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 influences subsequent search behavior. These results support the use of explicit context--behavior--outcome memory as an active mechanism for controlling population-based search.
1 Introduction
Existing optimizers retain solutions, parameter values, or operator performance, but often do not represent the search conditions under which behavior succeeds or fails. CMDO stores context–behavior–outcome experience across multiple memory timescales and retrieves it to guide future search; it is evaluated on numerical benchmarks and photovoltaic estimation.
1 Introduction
- Existing methods use population dynamics, parameter adaptation, or past performance, while CMDO asks what search conditions made a behavior succeed or fail.The introduction frames this as the question of what an optimizer should remember.Source: §1 · pp. 1–2
- CMDO is evaluated on selected BBOB/COCO and CEC2017 problems and photovoltaic parameter estimation, against DE, CMA-ES, SHADE, GWO, HHO, and ORCA.The photovoltaic experiments use a two-diode model and measured current–voltage data.Source: §1 · p. 2
- CMDO represents experience as the relationship between search context, applied behavior, and observed outcome, rather than only retaining successful solutions or parameter values.Retrieved experience guides exploratory, directed, and local search behavior.Source: §1 · p. 2
- Its working, episodic, and consolidated memories use contextual retrieval, reinforcement, consolidation, forgetting, and continued experience acquisition.Both positive and negative experience are retained to reinforce useful behaviors and discourage unsuccessful ones in related states.Source: §1 · p. 2
2 Related Work and Motivation
Prior optimization methods retain successful parameters or aggregate operator performance, while episodic-memory research emphasizes context-sensitive retrieval and retention. CMDO applies these ideas by storing context-conditioned search experience, including negative evidence, to guide future optimization actions.
2.1 Memory and adaptation in optimization
- Unlike methods that adapt parameters or select operators from aggregate rewards, CMDO links each search behavior to the state in which it was applied.Relevant conditions include population diversity, search progress, stagnation, and proximity to promising regions.Source: §2.1 · p. 3
2.2 Episodic and context-dependent memory
- Episodic-control, replay, and memory-augmented sequence models illustrate context-sensitive retrieval and retention across tasks beyond numerical optimization.These approaches motivate deciding what to retain, when to retrieve it, and how it should affect current computation.Source: §2.2 · p. 3
2.3 From neurocognitive memory to search experience
- CMDO stores search state, applied behavior, and outcome across working, episodic, and consolidated memory, retrieving experiences by similarity to current context.Successful experience can strengthen a behavior, while failure in a similar condition can reduce support for repeating it.Source: §2.3 · pp. 3–4
3 Cognitive Memory-Driven Optimization Algorithm
CMDO represents search experience as context, recipe, and outcome, then uses relevant memories to guide search behavior. Its optimization loop records new outcomes across multiple memory timescales, including unsuccessful actions.
3.1 Overview
- CMDO stores a search experience linking context, recipe, and reward, retrieves relevant experience to guide the next action, and writes outcomes back into memory.It maintains a population and generates one new candidate at each search step.Source: §3.1 · p. 4
3.2 Search context
- Its eight context features describe population spread, focal-solution rank and position, evaluation progress, stagnation, recent positive rewards, and the fitness gap to the best solution.The context uses only information available during optimization, not knowledge of the true optimum.Source: §3.2 · pp. 4–5
3.3 Search recipe and candidate generation
- A search recipe specifies an operator, step scale, directional mixture, and search geometry, with behaviors spanning exploration, directed search, and local refinement.Geometry can restrict moves to a randomly selected subset of dimensions rather than the full space.Source: §3.3 · p. 5
3.4 Multi-timescale memory
- Working memory holds recent experiences, episodic memory retains a longer history, and consolidated memory preserves patterns repeatedly associated with useful outcomes.The memories store context, recipe, and outcome rather than candidate locations.Source: §3.4 · p. 5
3.5 Context-dependent retrieval
- CMDO retrieves memories by contextual similarity, reduces older memories' influence, and considers relevant records for recipe selection.A stored experience receives reuse credit only if its recipe is selected and executed.Source: §3.5 · p. 5
3.6 Learning from positive and negative experience
- Positive experience supports repeating recipes in related contexts, while negative experience reduces support for behavior that failed under comparable conditions.Negative evidence applies only to behaviorally similar recipes, including the same operator and geometry; fresh recipes preserve opportunities to discover new behaviors.Source: §3.6 · p. 6
3.7 Reward and memory update
- CMDO compares each candidate's objective value with its focal solution, records improvement or failure, and replaces the focal solution only when the candidate improves the objective.The resulting experience enters working and episodic memory, and repeatedly useful experience may be consolidated; recalled recipes are reconstructed from the current population rather than replaying old solutions.Source: §3.7 · p. 6
4 Experimental Results and Discussion
CMDO performs competitively, with results varying by problem, and memory recall changes which search behaviors are selected. The experiments also show unsuccessful experiences remain available to guide later decisions and extend the same memory architecture to photovoltaic parameter estimation.
4.1 Numerical optimization performance
- 1322.72 median error on CEC2017 F10 is lower than all six comparison methods; on BBOB, CMDO is competitive but its relative performance varies by problem.On BBOB F1, its median error is 1.45 × 10−3, lower than HHO, GWO, ORCA, CMA-ES, and SHADE.Source: §4.1 · p. 7
4.2 How memory changes the search
- Recalled recipes develop problem-dependent operator preferences: directed search dominates on four functions, while other functions retain more balanced or mixed distributions.Fresh recipes stay close to the predefined operator distribution, so the shifts reflect accumulated experience rather than a changed operator prior.Source: §4.2 · p. 7
4.3 Negative experience and memory persistence
- Unsuccessful experiences remain available as negative evidence; on CEC2017 F10, late-search memory contains substantially more negative than positive original experiences.Working, episodic, and consolidated memory preserve recent events, a broader history, and repeatedly useful patterns, respectively.Source: §4.3 · p. 8
4.4 Photovoltaic parameter estimation
- On the photovoltaic task, CMDO's median MAE is 0.005402 A, lower than HHO, GWO, ORCA, and SHADE, while DE and CMA-ES obtain lower median values.The same memory architecture and search mechanism are used without a photovoltaic-specific operator or memory rule.Source: §4.4 · pp. 8–9
4.5 Discussion
- The discussion attributes CMDO's contribution to context-dependent memory that reshapes search and uses both successful and unsuccessful experience; evaluation is limited to low-dimensional benchmarks and one measured photovoltaic curve.Future work is proposed for higher-dimensional, noisy, constrained, dynamic, and multi-objective settings, and for transfer across related tasks.Source: §4.5 · p. 9
5 Conclusion
CMDO represents experience through context, behavior, and outcome, using relevant past experience to shape new actions rather than replaying old candidate locations. Its search context uses information available during optimization, not knowledge of the true optimum.
5 Conclusion
- CMDO represents search experience through the relationship between context, behavior, and outcome, and retrieves relevant experience to shape new actions under similar conditions.Its conclusion characterizes memory as an active search-control mechanism and identifies richer contexts, broader settings, and cross-problem transfer as future work.Source: §5 · p. 10
A CMDO Algorithmic Details
- The algorithm operates in normalized [0, 1]^D space and maps candidates to physical problem bounds before objective evaluation.Initialization evaluations count toward the evaluation budget.Source: A CMDO Algorithmic Details · p. 12
A.1 Overall CMDO procedure
- At each step, CMDO selects a focal population member, retrieves experience, chooses a recalled or fresh recipe, constructs and evaluates a candidate, then records the outcome.Recalled recipes are reconstructed from the current population rather than replaying old candidate vectors.Source: §A.1 · p. 12
A.2 Search context
- The eight-dimensional context describes population spread, focal rank and distances, budget consumed, stagnation, recent positive-reward frequency, and fitness gap to the best.The search controller does not use the benchmark's true optimum.Source: §A.2 · pp. 12–13
A.3 Fresh recipes and search geometry
CMDO recipes specify an operator, scale, directional mixture, and search geometry, enabling different search moves. Retrieval uses relevant experiences while applying negative evidence only to behaviorally similar recipes.
A.3 Fresh recipes and search geometry
- A search recipe encodes the operator, step scale, directional mixture, and whether the move uses full-space or subspace geometry.The three behaviors are exploration, directed search, and local refinement.Source: §A.3 · p. 13
- Fresh recipes sample exploration, directed search, or local refinement with probabilities 0.5, 0.3, and 0.2, respectively.Subspace geometry changes selected coordinates while leaving the rest equal to the focal solution.Source: §A.3 · p. 13
A.4 Retrieval and negative evidence
- Retrieval filters records by relevance and considers the top-k; positive records can propose recipes, while negative records penalize matching operator-and-geometry behavior.If no recipe has positive support, CMDO uses a fresh recipe.Source: §A.4 · p. 14
A.5 Memory update and consolidation
Each search step records its context, executed recipe, reward, and displacement, whether the outcome is positive or negative. Reuse feedback and compatibility determine which experiences are consolidated into bounded memory stores.
A.5 Memory update and consolidation
- Every completed step writes an experience containing context, recipe, reward, step, and realized displacement to working and episodic memory, regardless of reward sign.Source: §A.5 · p. 14
- An original experience becomes eligible for consolidation after at least three selected reuses with a positive mean reuse return.Reuse feedback is assigned only to the selected source record.Source: §A.5 · p. 14
- Compatible records share operator and geometry and satisfy the stated context-distance threshold; they update a consolidated prototype, otherwise a new prototype is formed.Working memory holds the newest 24 experiences, episodic memory is bounded at 256 records, and consolidated memory at 24 prototypes.Source: §A.5 · p. 14
B.1 Benchmark protocol
The benchmark protocol evaluates selected BBOB/COCO and CEC2017 functions at specified dimensions and evaluation budgets, using instance 1 and seeds 15–17.
B.1 Benchmark protocol
- The reported panel uses BBOB/COCO F1, F8, and F15 at D = 5 with 500 evaluations, and CEC2017 F1, F4, and F10 at D = 10 with 1000 evaluations.Instance 1 and seeds 15–17 are used; initialization evaluations count toward each budget.Source: §B.1 · p. 15
C Additional Convergence Results
Additional convergence trajectories complement final-error tables by showing when improvements occur within the fixed evaluation budget. They also illustrate how CMDO’s search behavior evolves across the selected landscapes.
C Additional Convergence Results
- The additional trajectories show when improvements occur within the fixed evaluation budget, complementing the final-error tables.The figure includes the representative F10 example for completeness alongside the remaining recorded trajectories.Source: C Additional Convergence Results · p. 15
- These trajectories complement the final-error tables by showing when improvements occur within the fixed evaluation budget and illustrate the problem-dependent evolution of CMDO’s search behavior across the selected landscapes.Source: C Additional Convergence Results · p. 15
D Photovoltaic Model and Experimental Details
The photovoltaic experiment applies CMDO to a seven-parameter two-diode model using measured RTC France current–voltage data. Its protocol specifies parameter representation, common evaluation conditions, and memory-guided search steps.
D.1 Two-diode photovoltaic model
- The two-diode model is defined implicitly for measured voltage and current, and RMSE is calculated after optimization as an additional fit-quality measure.The parameter-bounds table covers the RTC France two-diode model.Source: §D.1 · pp. 15–16, Table 4 · p. 16
D.2 Parameter bounds and representation
- The seven-dimensional normalized search maps I01 and I02 logarithmically and the remaining parameters linearly into the physical search space.Table 4 summarizes the RTC France two-diode parameter bounds.Source: §D.2 · p. 16, Table 4 · p. 16
- The RTC France experiment uses 26 measured I–V observations at 33◦C and 1000 W/m2; each optimizer receives 2000 objective evaluations for seeds 101–110.Source: §D.2 · p. 16
- The algorithm initializes a normalized population and working, episodic, and consolidated memories before iterating until the evaluation budget is reached.The memory system is paired with reward history and a stagnation counter.Source: §D.2 · p. 17
- At each iteration, CMDO computes the search context, retrieves relevant records, selects a recalled recipe or fresh exploration, and constructs a direction from the current population.Search geometry is selected as full-space or subspace according to the recipe.Source: §D.2 · p. 17
- A candidate is mapped to the physical search space and replaces the focal solution if its objective is lower; executed recalled sources receive updated reuse statistics.CMDO also updates the global best and stagnation state, records new experience, consolidates repeatedly useful experience, and applies bounded-memory eviction.Source: §D.2 · p. 17