IoT-Based Reinforcement Learning Using Probabilistic Model for Determining Extensive Exploration through...

Area of Science:

• Computational Intelligence
• Machine Learning
• Internet of Things (IoT)

Background:

• Reinforcement learning (RL) often struggles to balance exploration and exploitation.
• Deep learning in RL primarily enhances generalization, neglecting exploration challenges.
• Integrating computational intelligence with IoT offers potential for advanced learning techniques.

Purpose of the Study:

• To propose a deep reinforcement algorithm enhancing exploration effectiveness using intelligent sensors and Bayesian methods.
• To extend Bayesian parameter posterior distribution computation to nonlinear models like artificial neural networks.
• To introduce the Bayesian Bootstrap Deep Q-Network (BBDQN) algorithm.

Main Methods:

• Development of a deep reinforcement algorithm integrating computational intelligence, intelligent sensors, and the Bayesian approach.
• Expansion of Bayesian linear regression techniques for posterior distribution computation in nonlinear models.
• Creation of the Bayesian Bootstrap Deep Q-Network (BBDQN) by combining bootstrapped Deep Q-Network (DQN) with Bayesian computation.

Main Results:

• The proposed BBDQN algorithm demonstrates superior exploration efficiency compared to DQN and bootstrapped DQN.
• Effectiveness validated in two scenarios specifically designed to present significant exploration challenges.
• The Bayesian approach successfully addresses the exploration-exploitation dilemma in deep reinforcement learning.

Conclusions:

• BBDQN offers a significant advancement in reinforcement learning exploration, particularly for IoT applications.
• The integration of Bayesian methods and deep learning provides a robust framework for complex learning tasks.
• The algorithm effectively tackles severe exploration problems, improving overall learning performance.