Artificial Intelligence-Reinforcement Learning

Reinforcement learning (RL) is the science of decision-making, and is an incredibly general framework for learning optimal behaviors in an environment, exactly like how human beings learn the actions that help them achieve a goal. In principle, a robust and excellent RL system should be great at learning anything. Merging this framework with the empirical power of deep learning can achieve very good state representations of very challenging tasks to solve real-world problems. Therefore, deep RL (DRL) is one of the closest things that look anything like artificial general intelligence (AGI). DRL has been studied in a variety of challenging action control and decision-making tasks. We have been working on novel techniques to improve the performance of RL systems. The following video shows one of our recent works published in International Joint Conference on Artificial Intelligence (IJCAI)'22 (Acceptance rate:15%). IJCAI is one of the most prestigous conferences in AI. We proposed a novel multi-constraint DRL method to solve the action shaking problem.

Quantum Machine Learning

Quantum computing and machine learning are both transformational technologies, and machine learning is likely to require quantum computing to achieve significant progress. Although machine learning produces functional applications with classical computers, it is limited by the computational capabilities of classical computers. Quantum computing can be used for the rapid training of machine learning models and to create optimized algorithms. An optimized and stable machine learning method provided by quantum computing can complete years of analysis in a short time and lead to advances in technology. Thanks to the computational advantages of quantum computing, quantum machine learning can help achieve results that are not possible to achieve with classical computers. We have been working on applying quantum computing in the field of machine learning to use the superposition of quantum states and the acceleration of quantum algorithms to solve the problem of huge data volume and slow training process in current algorithms.

Smart Environments

Smart environments link computers and other intelligent devices to everyday settings and tasks. Smart environments include smart homes, smart cities, smart manufacturing, etc. Smart environments advance an otherwise passive environment to become an active partner of its users, and have the potential to allow intelligent devices to engage and interact seamlessly with their immediate surroundings. Smart environments are an active area of research and constitute a promising source of innovation, including training intelligent systems using the massive data generated from smart environments.

Most of existing works on smart environments do not consider 3D scenes of physical environments, which limited the interactions between intelligent devices and their surroundings. Recent advances in neural radiance field (NeRF) using a fully-connected neural network can generate novel views of complex 3D scenes, based on a partial set of 2D images. We have been working on the recent advances of NeRF technologies to seamlessly integrate the 3D scenes with intelligent devices. The following video shows a 3D scene generated from 2D images that can be used to train an autonomous vehicle.

Applications of Artificial Intelligence

AI has a wide range of applications, including speech recognition, text analysis, robotics, healthcare, etc. Robotics will be critical in the era of AI, among which we mainly focus on the specific areas of autonomous vehicles. Generally, there are two major paradigms for robot operations: the traditional rule-based control method and the data-driven learning method. Control theories rely on human intelligence to design rules directly or indirectly in advance. Data-driven learning algorithms reduce or even avoid the dependence on human intelligence through heuristic learning. Recently, with the success of deep learning and RL, DRL has become a powerful learning framework for tackling control problems with high dimensional inputs . More and more people have focused on using the deep reinforcement learning (DRL) for robot manipulations. We have proposed several novel schemes to improve the performance of autonomous driving. The following videos show an autonomous vehicle and an robot arm using our algorithms.