An Efficient Parameter-Free Learning Automaton Scheme.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Reinforcement Learning

Background:

• Learning Automata (LA) are crucial for intelligent agents interacting with stochastic environments.
• Traditional LAs require extensive parameter tuning, which is time-consuming and computationally expensive.
• Parameter-free LAs are highly sought after to reduce training costs and improve efficiency.

Purpose of the Study:

• To propose an Efficient Parameter-Free Learning Automaton (EPFLA) that eliminates the need for environment-specific parameter tuning.
• To introduce a novel Separating Function (SF) that integrates frequentist and Bayesian inference for action evaluation and selection.
• To mathematically prove the ϵ-optimality of the proposed EPFLA scheme.

Main Methods:

• Developed an Efficient Parameter-Free Learning Automaton (EPFLA) utilizing a Separating Function (SF).
• The SF combines frequentist and Bayesian inference to assess action performance differences and guide exploration.
• Formal proof provided to guarantee the ϵ-optimality of the EPFLA.

Main Results:

• EPFLA demonstrates superior performance compared to traditional parameter-based learning automata.
• The proposed parameter-free approach significantly outperforms existing parameter-free methods.
• EPFLA achieves optimal behavior efficiently without requiring pre-training or parameter adjustments.

Conclusions:

• EPFLA offers an efficient and effective parameter-free solution for reinforcement learning tasks.
• The novel SF design enhances action selection and performance evaluation in unknown environments.
• This research advances the development of more adaptable and less computationally intensive learning automata.