Oraculum: a model for self-adaptive system optimization in smart environments

Abstract

This dissertation introduces Oraculum, a modular self-adaptive framework designed to support the monitoring, prediction, reasoning, and adaptation of distributed systems operating in smart environments. Many existing solutions treat these tasks as disconnected components, relying on static training phases, fixed adaptation logic, and reactive decision-making triggered only after system degradation is detected. Oraculum proposes an integrated approach in which monitored metrics are continuously collected and processed to generate predictions and select actions in advance of performance failures. The framework consists of three key components. SHiELD is a sensor data simulator that generates synthetic time-series data using ARIMA models and applies heuristic methods-such as filtering, aggregation, and compression-to simulate realistic variability and reduce processing overhead. OntOraculum is a semantic ontology that formalizes performance metrics into five categories and enables the system to classify and validate alerts through rule-based reasoning and SPARQL queries. The adaptation engine uses regression and classification models to forecast short-term metric behavior and integrates a reinforcement learning agent based on a Markov Decision Process (MDP), which receives contextual states and selects actions such as resource scaling, scheduling adjustment, or service reconfiguration. The RL engine also includes a retraining mechanism that periodically updates policies using new data. The entire architecture operates in a closed feedback loop, using predictions and inferred knowledge to support earlier and more informed decisions. The model includes automated pipelines for dataset creation, model training, hyperparameter tuning, and continuous learning, covering both predictive models and RL agents. Experimental validation was conducted in a containerized testbed with simulated load variation. Results were collected across multiple performance indicators, including CPU, memory, latency, and model accuracy. The contributions of this work are: (i) the proposal of an integrated framework that combines monitoring, forecasting, semantic validation, and adaptation; (ii) the development of SHiELD for synthetic data generation and heuristic preprocessing; (iii) the design of OntOraculum for metric classification and rule-based inference; (iv) the implementation of a prediction-based strategy for early alert generation to reduce adaptation delay; and (v) the modeling of an RL engine with configurable actions and scheduled policy retraining.