TLDR: Percepta is a lightweight Data Stream Processing (DSP) system designed for Edge Computing, specifically to support AI workloads, with a strong focus on Reinforcement Learning (RL). It addresses limitations of cloud-centric solutions by processing data near its source, reducing latency, and enhancing privacy. Key features include real-time reward function computation for RL, data storage for model retraining, data normalization and harmonization across heterogeneous protocols and sampling rates, and robust handling of missing data. Its modular architecture allows flexible deployment from edge devices to cloud environments, making it suitable for real-world Edge AI applications like energy management.
The world is increasingly filled with real-time data, thanks to the explosion of Internet of Things (IoT) devices. From smart homes to industrial sensors, data is constantly being generated. However, relying solely on cloud computing for processing all this data presents significant challenges, including delays (latency), limited network capacity (bandwidth), and privacy concerns. This is where Edge Computing comes in, bringing data processing closer to where the data is generated.
While Edge Computing helps, it still needs efficient ways to handle continuous streams of data, especially when Artificial Intelligence (AI) models are involved. Traditional Data Stream Processing (DSP) systems like Apache Spark or Storm often require substantial computing power, which isn’t always available on smaller edge devices.
This is where Percepta, a new lightweight Data Stream Processing system, steps in. Developed by researchers from INESC-TEC/Instituto Superior de Engenharia do Porto, Percepta is specifically designed to support AI workloads at the edge, with a particular focus on Reinforcement Learning (RL) algorithms. RL is a type of AI where an agent learns by interacting with its environment and receiving feedback (rewards) for its actions.
Percepta offers several specialized features that set it apart. It can compute reward functions directly from real-world interactions at the edge, which is crucial for RL models to learn and improve continuously. It also stores necessary data for future model retraining, ensuring that AI models can adapt over time. Beyond RL, Percepta handles real-time data preparation, including normalization and harmonization of data from various sources with different communication protocols (like MQTT, HTTP, AMQP) and reporting rates. It also robustly manages missing or incomplete data, a common issue in IoT environments, by detecting gaps and filling them in.
The architecture of Percepta is modular and designed for flexibility. It includes several key components:
Also Read:
- Greener Clouds: Federated Learning Predicts Data Center Energy Use While Protecting Privacy
- Collaborative AI for Smart Factories: Understanding Federated Split Learning for Resource-Constrained Robots
Percepta’s Core Components
Receivers: These are the entry points for data, adapting to different communication protocols from various devices.
Translators: They take raw data from Receivers, filter out irrelevant information, and structure it into a standardized format before sending it to an internal queue.
Accumulator: This component collects data from the queue for a specific processing environment.
Manager: At regular intervals, the Manager processes the collected data. It can aggregate data (e.g., calculating averages), predict missing values based on historical data, and identify relationships between different data streams. The processed data is then sent to the Predictor.
Predictor: This component routes the prepared data to the appropriate AI model, collects predictions, validates them, and computes rewards. It also logs all input data, decisions, and rewards for future analysis or retraining.
Encoder/Decoder: Since different AI models may require unique data formats, this component translates the standardized data into the model-specific format before inference and decodes the model’s decisions back into a common format.
Forwarders: These components are responsible for transmitting the AI model’s decisions to the relevant external systems or devices, ensuring they are formatted and sent correctly.
Percepta’s modular design allows it to integrate new data sources easily by simply implementing a dedicated Receiver and Translator. It also supports multiple independent processing environments, meaning it can serve a single local environment at the edge or multiple isolated environments in a cloud-based deployment.
The system’s flexibility extends to its deployment. While primarily designed for the edge, Percepta can run on standalone edge devices, in fog environments serving multiple nearby locations, or entirely in the cloud. This adaptability is facilitated by its containerization support using Docker and Kubernetes.
A real-world application of Percepta is within the OPEVA project, which aims to optimize the energy consumption of buildings using an AI model. In this scenario, Percepta manages the complex data flow from various sources like energy production, consumption, EV charging data, and external APIs (e.g., energy prices, weather). It handles the diverse protocols and formats, corrects anomalies, imputes missing values, encodes data for the AI model, calculates rewards, stores data, and forwards control commands to devices. This demonstrates Percepta’s potential to bridge the gap between AI simulations and real-world energy management systems.
In conclusion, Percepta offers a robust and lightweight solution for high-performance data stream processing at the edge, specifically tailored for AI workloads, particularly Reinforcement Learning. Its capabilities in data harmonization, error correction, and native RL support make it a promising tool for the growing demands of Edge AI. For more details, you can refer to the original research paper.
