Optimizing Reinforcement Learning Using a Generative Action-Translator Transformer
-
Published:2024-01-16
Issue:1
Volume:17
Page:37
-
ISSN:1999-4893
-
Container-title:Algorithms
-
language:en
-
Short-container-title:Algorithms
Author:
Li Jiaming1, Xie Ning1, Zhao Tingting23
Affiliation:
1. Center for Future Media, School of Computer Science and Engineering, and Yibin Park, University of Electronic Science and Technology of China, Chengdu 611731, China 2. College of Artificial Intelligence, Tianjin University of Science and Technology, Tianjin 300457, China 3. RIKEN Center for Advanced Intelligence Project (AIP), Tokyo 103-0027, Japan
Abstract
In recent years, with the rapid advancements in Natural Language Processing (NLP) technologies, large models have become widespread. Traditional reinforcement learning algorithms have also started experimenting with language models to optimize training. However, they still fundamentally rely on the Markov Decision Process (MDP) for reinforcement learning, and do not fully exploit the advantages of language models for dealing with long sequences of problems. The Decision Transformer (DT) introduced in 2021 is the initial effort to completely transform the reinforcement learning problem into a challenge within the NLP domain. It attempts to use text generation techniques to create reinforcement learning trajectories, addressing the issue of finding optimal trajectories. However, the article places the training trajectory data of reinforcement learning directly into a basic language model for training. Its aim is to predict the entire trajectory, encompassing state and reward information. This approach deviates from the reinforcement learning training objective of finding the optimal action. Furthermore, it generates redundant information in the output, impacting the final training effectiveness of the agent. This paper proposes a more reasonable network model structure, the Action-Translator Transformer (ATT), to predict only the next action of the agent. This makes the language model more interpretable for the reinforcement learning problem. We test our model in simulated gaming scenarios and compare it with current mainstream methods in the offline reinforcement learning field. Based on the presented experimental results, our model demonstrates superior performance. We hope that introducing this model will inspire new ideas and solutions for combining language models and reinforcement learning, providing fresh perspectives for offline reinforcement learning research.
Funder
National Key R&D Program of China Chengdu Science and Technology Project National Natural Science Foundation of China Intelligent Terminal Key Laboratory of SiChuan Province
Subject
Computational Mathematics,Computational Theory and Mathematics,Numerical Analysis,Theoretical Computer Science
Reference27 articles.
1. Lu, Y., and Li, W. (2022). Techniques and Paradigms in Modern Game AI Systems. Algorithms, 15. 2. Prudencio, R.F., Maximo, M.R.O.A., and Colombini, E.L. (2023). A survey on offline reinforcement learning: Taxonomy, review, and open problems. IEEE Trans. Neural Netw. Learn. Syst. 3. Devlin, J., Chang, M.W., Lee, K., and Toutanova, K. (2019, January 2–7). Bert: Pre-training of deep bidirectional transformers for language understanding. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Minneapolis, MN, USA. 4. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., and Askell, A. (2020, January 6–12). Language models are few-shot learners. Proceedings of the Advances in Neural Information Processing Systems 33, Online. 5. OpenAI (2023). GPT-4 Technical Report. arXiv.
|
|