Dynamic orchestration of transformer models for sentiment analysis on social networks: Optimized performance and sustainable computing

This work proposes an innovative dynamic orchestration framework for sentiment analysis of texts from social media platforms, addressing the limitations of static models in resource allocation and performance optimization. Static or heuristic orchestration systems struggle to manage the variability and heterogeneity of loads, causing inefficiencies and overconsumption of resources. This work proposes a dynamic orchestration framework specifically designed for large-scale sentiment analysis, formulated as a multi-objective problem simultaneously integrating probabilistic latency, temporal variance, task completion rate, energy consumption and assignment constraints. The architecture is based on an empirically calibrated reinforcement learning agent enriched with a load prediction mechanism, able to jointly select the appropriate Transformer model and the optimal execution resource. Experiments conducted on realistic scenarios show that this approach goes beyond static and heuristic methods, ensuring better stability, faster convergence and increased adaptability to distributed environments. Results confirm the interest of dynamic orchestration to reconcile predictive performance, operational efficiency and resilience to load variations, while paving the way for the future integration of advanced predictors and criteria related to computational sustainability.