The institute for Artificial Intelligence and Cognitive Engineering (ALICE) focuses on AI, machine learning, pattern recognition, cognitive modelling and multi-agent systems since 2001. The institute has about 40 researchers in these areas. The University of Groningen (Rijksuniversiteit Groningen) has over 5000 researchers and 24000 students. It is a top-100 ranked university in several global rankings. The research group that will be involved in the Mantis project is the APS group: Autonomous Perceptive Systems.
Relevant Expertise for the project:
The APS group of ALICE has a unique expertise in machine learning, pattern recognition and big data mining. Apart from innovations at the algorithmic level of reinforcement learning and support-vector machines, we have the world’s first 24/7 learning system, that involves elicitation of user ’labels’ for raw data over internet and a continuous, autonomous learning process. The amount of data runs in the hundreds of terabytes, the number of instances runs in the hundreds of millions of data objects, many of them described with high (5000) or extremely high (50000) dimensions (data columns). Our unique configuration is able to exploit the strong points of both the human and machine components of a learning system. The human input consists of labelling detected unknown patterns, this effort must be minimised. The competence of the machine is the ability to create a usable agglomeration of patterns from a massive data collection. Good algorithms will be able to surf on Moore’s law, i.e., their performance will improve as computers become faster. An example application of this approach is our Monk search engine for historical handwritten manuscript collections. In May 2014, this system was able to learn to read historical Chinese texts within two weeks, with limited human efforts. We firmly believe in the use of optimising closed loops (neocybernetics) in order to obtain effective, resilient and autonomous systems.
Role in the project:
We will focus on a real-life maintenance problem involving vision and other real-life sensing, where the problem consists of reducing MTBF for incidents which in the current practice can only be solved by human operator intervention. Using data collection and user-based (’ipad’-based) annotation of problem conditions during the industrial process, a continuous learning cycle is initiated. In the first stage of the project we will focus on forecasting and detection of the problem, using a basic signalling method. The sensing component ideally focuses on not one but on several possible problem classes at a sensor location. In the second stage of the project, we will try to actually close the loop by means of intervention using a modern flexible and cost effective robot systems for solving general problems, when necessary. Actuators may consist of simple pressured air or magnetic methods for, e.g., obstacle removal. The problem-solving system will gradually develop a repertoire of problem conditions and a probability distribution for corrective actions and their computed applicability (utility values) given the current context. The proposed solution entails a loose coupling between the production system itself and the problem solver but a tight description of problem conditions and corrective actions. We will provide solutions that allow for convenient reconfiguration and reuse of learned material. Our direct partner is Philips, who will provide the (warehouse) process data. RUG will implement the pattern matching method and learning cycle. We will make additional computing resources and storage available on the massive Groningen high-performance computing cluster