An Empirical Investigation of Transfer Effects for Reinforcement Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Reinforcement Learning

Background:

• Reinforcement learning (RL) models for sorting tasks are known for lengthy training times, even with small datasets.
• Investigating methods to optimize RL training efficiency is crucial for practical applications.

Purpose of the Study:

• To evaluate the effectiveness of transfer learning in accelerating the training process of reinforcement learning algorithms.
• To compare the performance of transfer and non-transfer Q-learning methods for the sorting problem.

Main Methods:

• Q-learning was employed as the base reinforcement learning algorithm.
• The sorting problem served as a case study to assess training efficiency and resource utilization.
• Performance was measured by comparing total training steps and 'brain capacity' (updated Q-values).

Main Results:

• The difference in total training steps between transfer and non-transfer methods diminished as the size of the dataset (numbers to be sorted) increased.
• Both transfer and non-transfer reinforcement learning approaches exhibited similar brain capacities when reaching comparable training milestones.

Conclusions:

• Transfer learning offers potential benefits in reducing training time for reinforcement learning on sorting problems.
• The efficiency gains of transfer learning may vary depending on dataset size, with diminishing differences for larger datasets.