[2O6-OS-16a] OS-16

[2O6-OS-16a-03] Using HyperNetworks with reinforcement learning

〇Chika Sawano¹ (1. The Open University of Japan)

Keywords:Reinforcement learning, HyperNetworks, Deep learning

Deep reinforcement learning has been attracting attention and expectation, including in learning LLM (Large Language Models), but issues have accures such as "learning takes time", "complex and difficult to implement," and "difficult to search for good networks" .In such a background,I focused on HyperNetworks, which has advantages such as efficient search for good networks, improved flexibility, knowledge sharing, and reduced number of parameters. method to generate a large main network, called the target network.In this study, I compared the learning status of experiments in which HyperNetworks and normal networks.In the experiments, learning was attempted by changing the learning rate and batch size. In many experimental patterns, HyperNetworks outperformed the normal network.