MONAQ: Multi-Objective Neural Architecture Querying for Time-Series Analysis on Resource-Constrained Devices

TL;DR

MONAQ tackles time-series analysis on resource-constrained devices by reframing NAS as Multi-Objective Neural Architecture Querying (NAQ) powered by large language models. It introduces a multimodal query generation module to enrich LLM understanding of time-series data and a multi-agent LLM search to explore deployable architectures without runtime training feedback. Through extensive experiments on 15 datasets, MONAQ-discovered models outperform handcrafted baselines and NAS baselines while achieving smaller, faster models, demonstrating strong accuracy-efficiency trade-offs. The framework advances edge AI usability by enabling natural-language driven design and open-ended search spaces, with practical implications for on-device sensing across health, environment, and activity recognition domains.

Abstract

The growing use of smartphones and IoT devices necessitates efficient time-series analysis on resource-constrained hardware, which is critical for sensing applications such as human activity recognition and air quality prediction. Recent efforts in hardware-aware neural architecture search (NAS) automate architecture discovery for specific platforms; however, none focus on general time-series analysis with edge deployment. Leveraging the problem-solving and reasoning capabilities of large language models (LLM), we propose MONAQ, a novel framework that reformulates NAS into Multi-Objective Neural Architecture Querying tasks. MONAQ is equipped with multimodal query generation for processing multimodal time-series inputs and hardware constraints, alongside an LLM agent-based multi-objective search to achieve deployment-ready models via code generation. By integrating numerical data, time-series images, and textual descriptions, MONAQ improves an LLM's understanding of time-series data. Experiments on fifteen datasets demonstrate that MONAQ-discovered models outperform both handcrafted models and NAS baselines while being more efficient.

Figures (8)

• Figure 1: Comparison between (a) existing LLM-based NAS and (b) our proposed MONAQ framework.
• Figure 2: Overall procedure of our framework. MONAQ first receives a user prompt and a time series with descriptions. It then generates time-series images and processes all required information through the multimodal query generation module (§\ref{['section:mmts']}) to create an organized multimodal query. This query is subsequently shared across different specialized agents within the LLM agent-based multi-objective search module (§\ref{['section:agent_search']}). Once all agents successfully complete their tasks, the final model is returned to the user.
• Figure 3: Examples of representative time series images containing two-channel ECG signals.
• Figure 4: A complete example of multimodal query generation results, showing data and modeling aspects.
• Figure 5: Performance comparison of our MONAQ and the baselines in average accuracy, inference latency, and model complexity (size) for classification tasks.