A neuro-fuzzy approach for efficient adaptation in distributed IoT systems: enhancing concurrency and computational efficiency

Abstract

This paper presents a novel neuro-fuzzy approach designed to enhance the efficiency of self-adaptive systems, particularly within distributed Internet of Things (IoT) environments. The core challenge addressed is the overwhelming size of the adaptation space—the vast set of potential system configurations and actions a self-adaptive system must evaluate to respond to changing conditions. Searching this space is computationally expensive and slows down real-time decision-making, which is critical in dynamic, resource-constrained IoT networks. To solve this, the authors propose integrating an Adaptive Neuro-Fuzzy Inference System (ANFIS) into the standard MAPE-K (Monitor-Analyze-Plan-Execute with Knowledge) feedback loop of a self-adaptive system. This neuro-fuzzy network acts as an intelligent filter within the machine learning component, rapidly evaluating and pruning irrelevant adaptation options before detailed analysis, thereby dramatically reducing the search space.

The methodology's strength lies in its hybrid architecture, which combines the interpretability of fuzzy logic with the learning capabilities of neural networks. The system uses Gaussian membership functions to fuzzify qualitative environmental inputs (like temperature and traffic levels) and a five-layer ANFIS structure to learn and apply adaptive fuzzy rules. This allows the system to handle uncertainty and dynamically adapt to new conditions without extensive human intervention. The framework was implemented and rigorously tested using the DeltaIoT simulator, modeling a network of 15 sensors in scenarios such as smart campus monitoring.

Key experimental results demonstrate the significant advantages of the proposed neuro-fuzzy approach (ANFIS-AS). Compared to a baseline self-adaptive system (AS) and one using a conventional artificial neural network (AN-AS), the ANFIS-AS model achieved superior performance across critical metrics. It substantially reduced the average number of adaptation options analyzed, leading to lower computational overhead. Consequently, the system exhibited reduced packet loss, lower communication delay, and most notably, significantly improved energy efficiency—a vital factor for battery-operated IoT devices. The approach also proved highly effective in meeting complex adaptation goals, including thresholds, set-points, and optimizations, often outperforming state-of-the-art methods like DLASeR and ML4EAS in coverage and ranking accuracy.

In conclusion, this research successfully delivers a scalable and efficient solution for managing the complexity of self-adaptation in distributed IoT systems. By intelligently reducing the adaptation space, the neuro-fuzzy framework enables faster, more energy-efficient, and robust decision-making. The work provides a practical blueprint for enhancing the autonomy and performance of IoT applications in smart cities, industrial monitoring, and other domains where systems must continuously and efficiently adapt to concurrent, real-world data streams.