Reinforcement learning in random neural networks for cascaded decisions

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Biophysical neural networks transmit signals as voltage spikes, a dynamic more closely modeled by Random Neural Networks (RNNs).
• Optimizing sequential decisions to minimize total cost is a key challenge in artificial intelligence and control systems.

Purpose of the Study:

• To propose a reinforcement learning strategy for optimizing cascaded decisions in systems modeled by Random Artificial Networks (RANs).
• To analyze the performance of this strategy, particularly the influence of the reinforcement function, on a maze learning problem.

Main Methods:

• Modeling the system using Random Artificial Networks (RANs).
• Developing a weight update rule and a reinforcement function tailored for sequential decision-making.
• Evaluating the strategy's effectiveness through simulations on the maze learning problem.

Main Results:

• The performance of the reinforcement learning strategy is significantly influenced by the choice of the reinforcement function.
• Satisfactory results were obtained when the reinforcement function incorporated the recency effect, prioritizing recent experiences.

Conclusions:

• The proposed reinforcement learning strategy, utilizing RANs, offers a viable approach for optimizing sequential decisions.
• The recency effect in the reinforcement function is crucial for enhancing the learning efficiency and performance of the system.