Make intelligent technology ubiquitous

Most of the computational problems required to achieve AI have been theoretically proved to be NP (non-deterministic polynomial complexity) problems. For a real-world problem with moderate size, it is impossible to get the optimal solution in the acceptable time simply by piling up the computational resources or improving the speed of the computer (such as by using high-performance computers). On the other hand, human beings today, after hundreds of millions of years of evolution, seem to have much more “innate” intelligence. For example, they have abilities such as performing cognition, analysis and decision making in a relatively short period of time. Therefore, since Turing’s proposal of an “intelligent machine”, the original concept of AI has been to simulate the evolution process by computation and designing intelligent algorithms has been the most important process to building an intelligent machine.

Professor Yao’s team aims to explore the basic principles of evolutionary computation and design novel algorithms that are more intelligent than the existing ones. Currently, the team is focusing on various types of intelligent needs for diverse real-world applications, with the hope of studying and designing more efficient intelligent algorithms. One of their mid-term goals is to investigate evolutionary computation approaches for building artificial general intelligence, for example a common intelligent system that is capable of addressing heterogeneous intelligent tasks in a large variety of sectors or applications. This would allow AI to integrate into people’s daily life in a ubiquitous way, the same way that the Internet and cloud computing became mainstream.

Realizing two key abilities of human intelligence by evolutionary computation

Professor Yao’s team mainly focuses on realizing two of the most common abilities of human intelligence by evolutionary computation, namely (1) (data) analysis and learning, and (2) decision-making and optimization.

For intelligent analysis and learning, Professor Yao has pioneered the study of a neural network learning algorithm based on evolutionary computation (evolutionary neural network), and proposed a new neural network learning algorithm. This algorithm provides an important means to solve a number of bottleneck problems (such as network structure design, multimodality and multi-saddle point problems in the training of network weights, etc.) that have troubled neural networks as well as deep learning for a long time. Based on this, Professor Yao combined evolutionary computation with ensemble learning, and put forward the idea of an evolutionary neural network ensemble, which overcomes the limitations of a single neural network and can further boost the performance of a trained neural network. His publications on this topic have won the 2001 IEEE Donald G. Fink Paper Award (the only one in the world in that year) and the 2011 IEEE Transactions on Neural Networks Outstanding Paper Award (the only one in the world in that year).

In terms of decision-making and optimization, a lot of results have been achieved in the analysis and design of evolutionary algorithms for such complex optimization problems, such as multi-peak optimization, constraint optimization, dynamic optimization and multi-objective optimization. For example, as one of the first researchers in the world who rigorously analyzed the time complexity of the evolutionary algorithm with population size N (N> 2), Prof. Yao proposed the drift analysis technique for time complexity analysis, which bridges the gap between theoretical analysis and practice of evolutionary algorithms. Besides, his proposal of Cauchy mutation operator and its general form – Lèvy mutation operator enriches the optional operator library of evolutionary algorithm and gives the evolutionary algorithm the ability to adaptively choose search operators, which further forms the fundamental framework of a multi-strategy adaptive evolutionary algorithm. The proposal of a stochastic ranking technique overcomes several difficulties in constrained optimization, such as the difficulty to determine the weight of a penalty function, or to normalize the penalty and objective functions to the same scale. Most of these techniques have become standard components for modern evolutionary algorithms, in particular, the drift analysis technique is known as a breakthrough in theory of evolutionary computation.

Based on their fundamental research on evolutionary algorithms, Professor Yao’s team also carried out a large number of applied research projects in software engineering, industrial design, intelligent transportation and smart logistics. For example, together with Professor M. Harman of University College London (UCL) they proposed the “Search-based Software Engineering”, a cross-domain research direction in evolutionary computation and software engineering fields. At the same time, they also proposed a series of efficient intelligent algorithms for important problems in software engineering fields such as software module clustering, software development process planning, software defect prediction, software test case generation and testing resource allocation. With regards to engineering designs, they have cooperated with Japenese automotive manufacturer Honda and proposed intelligent design algorithms to solve problems in automobile design. In June 2017, Shenzhen Science and Technology Innovation Committee approved Professor Yao’s application to set up the Shenzhen Key Laboratory of Computational Intelligence, aiming to make new breakthroughs in low-cost computational intelligence, autonomous evolution technology of computational intelligence and reliability of intelligent computing system.

Evolutionary intelligence is facing unprecedented opportunities

In the era of AI2.0, evolutionary intelligent systems such as intelligent systems built by means of evolutionary computing, face unprecedented opportunities. On one hand, the research content of the term “evolutionary computation” will be greatly enriched by the emergence of evolvable terminal hardware and devices as well as the increasing popularity of Internet technologies. It is gradually becoming possible to build human-machine collective evolutionary systems that cover a variety of heterogeneous terminal devices and even human intelligence, for key fields like smart cities, intelligent manufacturing and intelligent medical care. On the other hand, the development of high-performance and cloud computing technology enable us to observe and adjust the trend of evolution at a much lower cost than before, paving the way to new foundations for novel evolutionary intelligent systems. In the coming era, the continuous progress made by Professor Yao’s team is expected to bring revolutionary development of intelligent manufacturing and equipment, smart logistics, smart cities, human-machine collaboration and interaction and other related application fields.

Contributed by: Research Group of Professor Xin Yao