Reinforcement learning-driven feature selection enhanced by an evolutionary approach tuning for criminal suspect...

Area of Science:

• Computer Science
• Artificial Intelligence
• Forensic Science

Background:

• Accurate criminal suspect identification is vital for justice and crime deterrence.
• Convolutional Neural Networks (CNNs) are common but face challenges in feature selection, class imbalance, and hyperparameter tuning.
• These limitations reduce the effectiveness of conventional CNN-based suspect identification methods.

Purpose of the Study:

• To develop a novel strategy for criminal suspect identification that overcomes limitations of conventional methods.
• To enhance the effectiveness of CNNs by integrating reinforcement learning (RL) and differential evolution (DE) algorithms.
• To improve feature selection, address class imbalance, and optimize hyperparameters in suspect detection.

Main Methods:

• Implemented an off-policy proximal policy optimization (Off-policy PPO) algorithm for reinforcement learning to address feature selection and class imbalance.
• Utilized a differential evolution (DE) algorithm, enhanced with k-means clustering for mutation strategy, to tune hyperparameters.
• Employed Off-policy PPO for its reduced data requirements and increased efficiency in costly data collection settings.

Main Results:

• The proposed methodology achieved high F-measures across four datasets: CelebA (89.409%), LFW (91.152%), CASIA-WebFace (92.184%), and VGGFace2 (92.202%).
• The Off-policy PPO dynamically tuned approach consistently surpassed conventional static methods in criminal suspect detection.
• The enhanced DE algorithm effectively identified key clusters through its novel mutation strategy.

Conclusions:

• The developed approach significantly outperforms existing methods in criminal suspect identification.
• This research advances early suspect detection capabilities and enhances investigative strategies.
• The integration of RL and DE offers a robust solution for complex identification tasks with limited or imbalanced data.