Evolutionary Design of Convolutional Neural Networks for Human Activity Recognition in Sensor-Rich Environments.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Human activity recognition is crucial for context-aware systems, driven by diverse sensor data.
• Current methods often rely on supervised machine learning, requiring manual feature engineering and model design.

Purpose of the Study:

• To propose a neuroevolutionary approach for human activity recognition in heterogeneous sensor networks.
• To optimize convolutional neural network (CNN) topology using evolutionary algorithms for improved classification F1 score.

Main Methods:

• Utilized deep learning, specifically CNNs, for activity recognition on the OPPORTUNITY dataset.
• Employed an evolutionary algorithm to automatically design optimal CNN topologies, avoiding manual design.
• Investigated the performance of model ensembles (committees) from the evolutionary process.

Main Results:

• The proposed neuroevolutionary system achieved high accuracy in recognizing human activities within sensor network environments.
• The automated topology design successfully maximized the classification F1 score.
• Ensemble models demonstrated robust performance on unseen sensor data.

Conclusions:

• The neuroevolutionary approach effectively identifies optimal deep learning models for human activity recognition.
• This method surpasses previous state-of-the-art results, demonstrating the utility of automated architecture search.
• The system offers a systematic way to improve upon manually designed neural network architectures.