R-POPTVR: a novel reinforcement-based POPTVR fuzzy neural network for pattern classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Intelligence

Background:

• Fuzzy neural networks (FNNs) typically learn from static data, lacking environmental interaction.
• Interactive problems necessitate agents learning from experience, aligning with reinforcement learning (RL) principles.

Purpose of the Study:

• To develop novel clustering algorithms for fuzzy neural networks based on the reinforcement learning paradigm.
• To integrate these RL-based clustering techniques into a fuzzy neural network architecture for improved performance.

Main Methods:

• Three reinforcement learning-based clustering algorithms were developed: REINFORCE clustering technique I (RCT-I), REINFORCE clustering technique II (RCT-II), and episodic REINFORCE clustering technique (ERCT).
• These algorithms were integrated with the pseudo-outer product truth value restriction (POPTVR) fuzzy neural network, creating RPOPTVR-I, RPOPTVR-II, and ERPOPTVR.
• Benchmarking was conducted using the Iris, Phoneme, and Spiral datasets.

Main Results:

• The RPOPTVR models demonstrated superior classification results compared to the original and modified POPTVR on the Iris and Phoneme datasets.
• RPOPTVR-II achieved a performance improvement of at least 5.8% over other methods on the Spiral dataset.
• The integration of RL-based clustering showcased a trial-and-error search characteristic, leading to enhanced qualitative performance.

Conclusions:

• Reinforcement learning-based clustering significantly improves the performance of fuzzy neural networks in classification tasks.
• The proposed RPOPTVR models offer a more adaptive and effective approach to clustering and classification, especially in interactive environments.
• These findings highlight the potential of combining reinforcement learning with fuzzy neural networks for complex problem-solving.