Modified Triplet-Average Deep Deterministic Policy Gradient for interpretable neuro-fuzzy deep reinforcement learning

In order to find the control rules of the nonlinear system from the learned data, it is necessary to interpret the learned policy in Deep Reinforcement Learning (DRL). This paper presents a novel interpretable Neuro-Fuzzy (NF) inference system based on Modified Triplet-Average Deep Deterministic Pol...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Journal of the Franklin Institute Jg. 362; H. 7
Hauptverfasser: Nguyen, Tuan-Linh, Thin, Nguyen Van, Lee, Sangmoon
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Elsevier Inc 01.05.2025
Schlagworte:
Interpretable neuro-fuzzy controller
Inverted pendulum
Reinforcement learning
Twin-delay
Two-phase training
Twin-delay
Inverted pendulum
Two-phase training
Reinforcement learning
Interpretable neuro-fuzzy controller
ISSN:0016-0032
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:In order to find the control rules of the nonlinear system from the learned data, it is necessary to interpret the learned policy in Deep Reinforcement Learning (DRL). This paper presents a novel interpretable Neuro-Fuzzy (NF) inference system based on Modified Triplet-Average Deep Deterministic Policy Gradient (MTADD) reinforcement learning algorithm with a two-phased training method. The first phase involves exploring and initiating the T-S fuzzy system rule and premise parameter. The second step is the deep reinforcement learning of the NF policy network, which uses a Modified Triplet-Average Deep Deterministic policy gradient algorithm. The experiment results demonstrate that the proposed approach decreases the training time, enhances the control performance, and increases the interpretability of NF DRL. •A novel Neuro-Fuzzy (NF) inference system is presented with a modified deep RL algorithm.•The new two-phase method for the neuro-fuzzy RL model reduces the training time.•Experiment results verify the effectiveness of the proposed approach.
ISSN:0016-0032
DOI:10.1016/j.jfranklin.2025.107653