Generative AI

3Reinforcement Learning in Generative AI

Kanthavel R.¹, Adline Freeda R.² and Dhaya R.^1*

¹School of Electrical and Communication Engineering, PNG University of Technology, Lae, Papua New Guinea

²Dept. of Information Technology, KCG College of Technology, Chennai, India

Abstract

For the recent generation, generative artificial intelligence (AI) has emerged as a potentially fascinating advancement in modern trends. In a similar period of time, reinforcement learning (RL) has evolved as a promising methodology across various machine learning applications. The machine learning community has long considered deep generative AI as a crucial topic in various applications such as computer vision and text generation. Maximum likelihood estimation (MLE) trains generative models to capture and approximate target data distributions by minimizing divergence between the model and target distributions. This formulation can effectively summarize the objective of generating tasks; however, it neglects some qualities that a generative model should possess according to an end user’s expectations. Different aspects of reinforcement learning have contributed to its flexibility to incorporate human bias, such as adversarial learning, hand-designed rules, and learned reward models for building preferment models. Reinforcement learning has become a popular and competitive option in the world of AI. It allows us to inject new training signals and create fresh objectives that take advantage of novel ...