A team of researchers from IIT Dharwad and TCS Research has proposed an innovative framework for time series analysis called Agentic Retrieval-Augmented Generation (RAG). This modular approach utilizes a hierarchical, multi-agent architecture to enhance flexibility and accuracy in handling complex time series tasks.
The framework consists of a master agent that orchestrates specialized sub-agents, each fine-tuned with small-scale pretrained language models (SLMs) for specific time series tasks like forecasting or anomaly detection. These sub-agents retrieve relevant prompts from prompt pools, which store historical patterns, enabling better predictions on new data.
This dynamic prompting mechanism overcomes the limitations of traditional fixed-window methods by allowing the model to adapt to different trends and patterns in complex time series data. The framework also incorporates a two-tiered attention mechanism to handle long-range dependencies in time series data.
The Agentic RAG framework was evaluated across various time series tasks, including forecasting, classification, anomaly detection, and imputation. It consistently outperformed baselines in forecasting tasks, demonstrating superior predictive accuracy and robustness across all metrics.
This innovative approach addresses challenges like distribution shifts and fixed-length subsequences, making it a promising solution for time series analysis applications.
Source: https://www.marktechpost.com/2024/09/01/agentic-rag-a-hierarchical-multi-agent-framework-for-enhanced-time-series-analysis/