Skip to Content
  • 8.4
    Impact Factor
  • 12.7
    CiteScore
  • 19 days
    Submission to First Decision
  • 4 days
    Acceptance to Publication

Machine Learning and Knowledge Extraction

Machine Learning and Knowledge Extraction is an international, peer-reviewed, open access, monthly journal on machine learning and applications, see our video on YouTube explaining the MAKE journal concept. 

Get Alerted

Add your email address to receive forthcoming issues of this journal.

All Articles (1,000)

  • Article
  • Open Access

ANEMMSO: An Adaptive Neighborhood-Enhanced Mosquito Mating Swarm Optimizer for Global and Engineering Optimization

  • Jesús Águila-León,
  • Rachid Marzoug and
  • Fabian Alonso Lara-Vargas
  • + 1 author

Swarm-intelligence optimizers are frequently introduced through new metaphors while remaining orders of magnitude behind differential evolution (DE) methods on standard benchmarks. This work asks whether a swarm optimizer can reach the boundary of the DE tier by taking its metaphor beyond itself: every operator justified by function and validated by one-factor-out ablation. The resulting algorithm, ANEMMSO, combines a three-leader hierarchy with inertial momentum, unified greedy selection, a current-to-pbest precision operator with success-history adaptation, population size reduction, and an operator schedule latched at diversity collapse; the schedule can also be learned during the run itself, with no pre-training and no external data. On the official CEC2017 evaluator (28 functions, four dimensions, 30 independent runs, sixteen comparators including four CEC competition winners), no swarm method in the panel, nor the genetic algorithm, outperforms ANEMMSO under the Holm procedure. On unimodal functions, it matches the accuracy of the best method of the panel, with a median error ratio of 1.00 at three of the four dimensions, and it improves on its canonical hierarchical formulation by 1.85 orders of magnitude on hybrid functions. The five DE methods remain ahead, with NL-SHADE-RSP the closest by a small effect size; composition functions at high dimension remain a documented limitation. An attribution analysis explains the hybrid’s behavior: on hybrid functions, the swarm move contributes three quarters of the descent while executing about six percent of the moves, so the hybrid is not reducible to its DE component. These results establish that a swarm optimizer taken beyond its metaphor, with every operator functionally justified, can reach the DE frontier.

Flowchart of ANEMMSO: generation-level loop (left column) and per-individual update loop (right column). Gray blocks mark the landscape-driven adaptation steps. The circled letters A and B are on-page connectors that link the two loops, and X* and f* denote the best solution found and its fitness value.
  • Article
  • Open Access

Low-rank adaptation (LoRA) is widely used for parameter-efficient fine-tuning, yet uniform rank assignment ignores the varying contribution of layers to a downstream task. Adaptive methods such as AdaLoRA allocate ranks via gradient and singular-value importance, but they do not exploit the representation change induced by the task. We introduce FLowRA, a calibration-based rank-allocation framework that combines a lightweight flow-change descriptor (velocity, acceleration, and directional alignment of hidden representations) with gradient activity to solve a constrained integer-budget problem. FLowRA complements existing LoRA variants, requires only two user-configurable parameters, and supports multiple architectures through architecture-specific implementations. On GLUE, SuperGLUE, instruction-following, code, vision, and vision-language benchmarks, FLowRA consistently outperforms uniform LoRA, DoRA, and AdaLoRA under low trainable-parameter budgets, reducing the gap to full fine-tuning by 48.5% on GLUE relative to the strongest baseline. We compare FLowRA with AdaLoRA using sensitivity-score correlations and report component ablations to assess the contribution of flow-guided allocation.

FLowRA architecture.
  • Article
  • Open Access

Thermal regulation in convective drying processes constitutes a complex nonlinear system characterized by the dynamic interaction of heat and mass transfer phenomena that govern drying kinetics and energy efficiency. This study develops and implements embedded machine learning algorithms for thermal control in an Internet of Things (IoT)-enabled edge computing cyber-physical convective drying system for jicama (Pachyrhizus erosus), deployed on an ESP32-S3 microcontroller. Three control strategies were compared: a decision tree classifier, a feedforward artificial neural network (ANN), and a conventional proportional–integral–derivative (PID) controller. The models were trained using a multivariable state vector integrating thermal and psychrometric process variables. Experimental validation with a 5 kg fresh jicama load showed that the embedded learning controllers achieved shorter drying times than the conventional PID controller under the investigated conditions. The ANN achieved the shortest drying time (8.08 h), highest thermal efficiency (80%), lowest specific thermal energy supplied by the drying air (3112 kJ/kg), and highest effective moisture diffusivity (2.0 × 10−10 m2/s). The decision tree exhibited competitive energy performance (3454 kJ/kg), demonstrating its applicability as an alternative embedded strategy for thermal regulation. In contrast, the PID controller resulted in the longest drying time (14.50 h). These findings demonstrate the feasibility of embedded machine learning to enhance thermal stability and energy efficiency in cyber-physical convective drying systems.

Instrumented experimental convective drying system for real-time data acquisition.
  • Article
  • Open Access

Regime-Aware Portfolio Rebalancing Using Hidden Markov Models and Reinforcement Learning

  • Ignacio Megías-Comín,
  • Íñigo Gómez-Carvajal and
  • Manuel Sánchez-Montañés

Portfolio rebalancing in multi-asset funds is typically governed by fixed calendar rules or threshold triggers, ignoring the information content of prevailing market conditions. This paper proposes a framework that combines Hidden Markov Models (HMMs) and Q-learning to learn regime-aware rebalancing policies for a benchmark-relative balanced portfolio. A Gaussian HMM is fitted on seven macro-financial features to produce interpretable and discrete market regime labels, which form the core of a structured 30-state Markov Decision Process encoding the current regime, existing portfolio allocation, and regime duration. Then, a Q-learning agent is trained to learn the best strategy (whether to rebalance or not) based on the market regime. This agent is benchmarked against a static 50/50 portfolio, a framing that reflects real-world private banking and wealth management practice. The agent delivers higher returns than the passive benchmark with positive information ratios in both in-sample and out-of-sample periods, executing a small number of rebalances. Unlike deep reinforcement learning approaches, Q-learning provides a fully inspectable and auditable policy. This is a critical requirement in private banking, where investment decisions must be explainable to clients and regulators. Comparison against a deterministic rule-based regime strategy confirms that the Q-learning component adds measurable value beyond the regime signal alone. We conclude that the proposed architecture offers a practical path toward systematic, low-turnover, and interpretable portfolio management, with clear potential for extension to richer state spaces and asset universes.

Cumulative total returns normalized to 100 at the start of the overlapping period (2010–2025) for the equity pair (left) and fixed income pair (right). The equity instruments show close co-movement over the evaluation period, with a cumulative total return difference of approximately 24 bps per year. The fixed income instruments show a larger divergence of approximately 60 bps per year, attributable primarily to differences in the maturity profile of their respective benchmark indices: VCIT tracks intermediate-term (5–10 year) corporate bonds, while the training fund tracks a broader investment grade universe; both instruments reflect the same asset class dynamics and directional regime behavior, and the divergence remains small relative to the allocation decisions being evaluated.

Highly Accessed Articles

News & Conferences

Latest Issues

Open for Submission

Journal Sections

XFacebookLinkedIn
Mach. Learn. Knowl. Extr. - ISSN 2504-4990