An adaptive inference model in mobile systems

Y.O. Haidukevych, A.Yu. Doroshenko

Abstract


The paper proposes and investigates a new model of adaptive distribution of the inference process (application of an ML model to obtain a prediction) between local and server-side computations for mobile intelligent forecasting systems. The goal of the proposed model is to overcome the fundamental contradiction between the requirement for high prediction accuracy (achieved through powerful server-side ML models) and the need to ensure low response time, autonomous operation, and energy efficiency on resource-constrained devices. The proposed model formalizes a dynamic mechanism for selecting the inference execution path (local TFLite, server-side microservice, or hybrid mode) based on the analysis of the execution context, including network connection quality, battery charge level, computational complexity of the request, and urgency of the result. The model is implemented within an architecture that combines a Flutter client with containerized microservices and is validated on a short-term meteorological forecasting task. Experimental results demonstrate that the proposed model reduces average response time by 35% compared to a purely server-based approach and decreases network traffic consumption by 60% compared to constant server usage, while maintaining prediction accuracy at the level of R² = 0.80–0.95 depending on the selected mode. The work has practical significance for the development of resource-efficient mobile applications in the fields of meteorology, environmental monitoring, and predictive analytics.

Problems in programming 2025; 4: 23-31


Keywords


adaptive inference; hybrid computing; mobile forecasting systems; on-device machine learning; network request optimization

References


Doroshenko A. Yu., Haidukevych Y. O., Haidukevych V. O., Zhyrenkov O. S. A Client Centric Technology Stack for Weather and Air Quality Forecasting. Programming Problems. – 2024.– No. 4. – pp. 18–22.

Haidukevych Y. O., Doroshenko A. Yu. On the Implementation of Meteorological Forecast Interfaces for Mobile Platforms. Programming Problems. – 2023.– No. 2. – pp. 14–19.

Haidukevych Y. O., Yatsenko O. A., Zhora D. V., Doroshenko A. Yu. Computer Program "Machine Learning and Visualization Software System for Meteorological Forecasts on Mobile Platforms (MeteoMob)". Copyright Registration Certificate No. 137634. Ukrainian National Office for Intellectual Property and Innovations, July 2, 2025.

Doroshenko A. Yu., Zhora D. V., Haidukevych V. O., Haidukevych Y. O., Yatsenko O. A. 24 Hour Ahead Nationwide Electrical Energy Consumption Forecasting. UkrProg-2024, CEUR Workshop Proceedings.– 2024.

Doroshenko A., Zhora D., Shpyg V., Haidukevych Y., Horvath R., Dimov I. Applying Machine Learning Techniques for Weather and Air Pollution Forecasting with Geographically Distributed Data. Proceedings of the International Conference on Artificial Intelligence, Computer, Data Sciences and Applications (ACDSA).– Antalya, Turkiye, 2025. pp. 1–6.

Doroshenko A.Y., Zhora D.V., Haidukevych V.O., Haidukevych Y.O., Yatsenko O.A. Predicting 24-Hour Nationwide Electrical Energy Consumption Based on Regression Techniques // CEUR Workshop Proceedings. 2024.

Doroshenko A.Y., Kushnirenko R.V., Yatsenko O.A. Designing a program for visualization of the Earth's surface using algebraic-algorithmic tools // Programming problems. – 2019, No. 2. – pp. 3–10.

Prusov V.A., Doroshenko A.Y. An efficient computational method for mesoscale weather forecasting // Dopovidi Natsionalnoi akademii nauk Ukrainy. – 2020, No. 3. – pp. 10–18.


Refbacks

  • There are currently no refbacks.