An Energy-Aware Approach to Design Self-Adaptive AI-based Applications on the Edge

Alessandro Tundo, Marco Mobilio, Shashikant Ilager, Ivona Brandic, Ezio Bartocci, Leonardo Mariani

Introduction

Deploying artificial intelligence at the "edge"—on localized devices such as Raspberry Pis or IoT sensors—presents a fundamental engineering conflict. While modern AI models are increasingly resource-intensive, edge devices typically operate under strict constraints, often relying on limited battery life or intermittent renewable energy sources like solar power. This paper introduces a novel energy-aware framework designed to bridge this gap by creating self-adaptive AI applications that dynamically balance performance and power consumption.

The Challenge of Edge AI

Engineers currently face two primary hurdles when designing edge-based AI systems:

  1. The Search Space Explosion: Configuring an AI pipeline involves numerous variables, including frame rates, image resolutions, and model architectures. The resulting configuration space can contain thousands of permutations, making exhaustive testing to find an optimal setup prohibitively expensive.
  2. Static Configuration Inefficiency: A "one size fits all" approach is inherently flawed. For instance, a pedestrian detection system optimized for a crowded city square at noon is unnecessarily power-hungry for a deserted street at 3 AM. Static applications either waste energy during low-activity periods or fail to provide necessary accuracy during critical events.

The Self-Adaptive Framework

The authors propose a four-step workflow to design a Self-Adaptive Application (SAA) that can respond to environmental changes in real-time:

1. State-Based Logic

Engineers first define a Finite State Machine (FSM) that represents different operational modes. These modes—such as power-saving, low-energy, high-accuracy, and high-rate—serve as the behavioral blueprint for the application.

2. Multi-Objective Optimization (MOOP)

To navigate the massive configuration space, the framework employs NSGA-II, a genetic meta-heuristic algorithm. This process identifies the "Pareto front"—a set of optimal configurations where no single objective (e.g., accuracy) can be improved without degrading another (e.g., energy efficiency).

3. Weighted Gray Relational Analysis (WGRA)

Once the Pareto front is established, the system must select a specific configuration for each FSM state. Using WGRA, engineers can assign weights to specific objectives based on the state’s priority. For example, in a power-saving mode, energy efficiency might be assigned 90% of the decision weight.

4. Runtime Implementation

The application is deployed using a MAPE loop (Monitor, Analyze, Plan, Execute). This loop continuously monitors environmental data—such as the number of detected pedestrians—and triggers state transitions to the most appropriate configuration.

Hardware Integration: To maximize energy savings, the framework incorporates hardware-level control. The team utilized a physical latching relay to completely power off external hardware accelerators, such as the Edge TPU, during low-intensity operational modes.

Experimental Evaluation

The approach was validated through a pedestrian detection case study. Accuracy testing was conducted using the ADL-Rundle-6 video from the Multiple Object Tracking (MOT) benchmark. To simulate real-world conditions, the researchers modeled "Weekday" and "Weekend" traffic patterns based on Manhattan pedestrian data. The study evaluated nine different pre-trained models from the COCO dataset, including various versions of SSD MobileNet and EfficientDet-Lite.

Key Results and Insights

Efficiency in Design

A significant finding was the efficiency of the meta-heuristic search. The NSGA-II algorithm required sampling only 10% of the total configuration space to identify solutions nearly identical in quality to those found by a near-exhaustive search that took two weeks to complete.

Performance Metrics

The self-adaptive system demonstrated substantial improvements over high-performance static baselines:

  • Energy Savings: The system achieved up to an 81% reduction in energy consumption during the "Weekend" scenario.
  • Accuracy Trade-offs: These massive energy savings were achieved with a minimal impact on performance, resulting in only a 2% to 6% drop in detection accuracy (mAP).
  • Throughput: The adaptive approach outperformed three out of four non-adaptive baselines in processing speed, with throughput gains ranging from 77% to 233% in frames per second.

Context-Aware Adaptation

The research highlights that the most significant energy gains do not stem from algorithm optimization alone, but from "downshifting." By reducing resolutions or powering down hardware when activity is low, the system preserves its energy budget for high-priority events. In simulations, the FSM successfully identified and adapted to the "three-peak" structure of typical workday traffic.

Conclusion and Future Work

This framework demonstrates that software-level adaptation, when paired with hardware-level power management, can significantly extend the viability of AI at the edge. While the current study relies on manual FSM design, future research aims to automate the synthesis of state machines using data-driven methods. Additionally, the authors plan to expand the framework to support clusters of edge devices and test the system against the unpredictable power fluctuations of real-world solar and battery cycles.

Summary updated

BibTeX
@inproceedings{tundo2023energy,
  author = {Alessandro Tundo and Marco Mobilio and Shashikant Ilager and Ivona Brandic and Ezio Bartocci and Leonardo Mariani},
  title = {{An Energy-Aware Approach to Design Self-Adaptive AI-based Applications on the Edge}},
  booktitle = {ASE},
  year = {2023},
  url = {https://openreview.net/forum?id=jsXFhEeGEC},
}
Figure