Deep Reinforcement Learning for Smart Steel Processing

顺序决策描述的是决策者在做出最终决策之前对一个过程进行连续观察的情况。随着近年来人工智能的发展，特别是深度学习(人工神经网络)的发展，在开发能够独立进行顺序决策的计算机代理方面取得了很大进展。在这个博士项目中，我们将开发先进的人工智能算法，用于智能钢铁加工中的顺序决策。博士项目是由EPSRC资助的计划的一部分，该计划由5家英国主要钢铁生产商(塔塔、Liberty、英国钢铁公司、Celsa、谢菲尔德锻造公司)和在该领域拥有专业知识的三所主要大学(斯旺西、华威和谢菲尔德)共同创建，以在钢铁创新领域提供学术领导。尽管钢铁行业产生大数据已有30多年，但到目前为止，生产效益有限。在这个博士项目中，我们将开发新的数据驱动技术，利用数据科学和机器学习的最新进展。最终，我们将为智能钢铁加工提供一个人工智能系统，能够自动化和优化某些仍然严重依赖人工干预的流程。我们将利用我们的行业合作提供的现有历史数据存储库，以及新一代传感器的可用性，这些传感器现在正在极端环境中取代传统的采样方法。我们还将开发“数字孪生”，这是一个基于模拟的环境，以帮助我们测试和开发新的强化学习算法。

Sequential decision making describes a situation where the decision maker makes successive observations of a process before a final decision is made. With recent advanced in artificial intelligence, and especially "deep learning" (artificial neural networks), much progress has been made in developing computer agents that are able to make sequential decision on their own. In this PhD project we will develop advanced artificial intelligence algorithms for sequential decision making for applications in smart steel processing.The PhD project is part of SUSTAIN, an EPSRC-funded programme co-created by the 5 major UK steel producers (Tata, Liberty, British Steel, Celsa, Sheffield Forgemasters) and the three principal Universities that have expertise in this area (Swansea, Warwick and Sheffield) to provide academic leadership in the field of steel innovation. Although the steel industry has generated "big data" for over 30 years, the production benefits have been limited so far. In this PhD project, we will develop novel data-driven techniques that leverage the latest advances in data science and machine learning. Ultimately, we will deliver an AI system for Smart Steel Processing able to automate and optimise certain processes that still rely heavily on manual intervention. We will exploit existing historical data repositories made available by our industrial collaboration and the availability of next-generation sensors that are now replacing traditional sampling methods in extreme environments. We will also develop a "digital twin", a simulation-based environment to help us test and develop novel reinforcement learning algorithms.