Pattern Recognition and Machine Intelligence

s of Invited Talks Interactive Granular Computing in Data Science Andrzej Skowron 2 1 Faculty of Mathematics, Computer Science and Mechanics, University of Warsaw, Poland skowron@mimuw.edu.pl 2 Systems Research Institute, Polish Academy of Sciences We discuss Interactive Granular Computing (IGrC) as the basis of a Data Science computing model. IGrC binds together and brings a synchronous cooperation among the following four basic concepts of Artificial Intelligence: language, reasoning, perception, and action. This, together with information granulation, helps agents to deal with many complex tasks of perceiving or transforming compound abstract and physical objects (e.g., in the context of complex spatio-temporal space). One should consider that in Data Science agents collecting data have control over the data acquisition, i.e., they are deciding say which data, using which sources, at what time, and why should be collected. Basic objects in IGrC are complex granules (c-granules or granules, for short). They are grounded in the physical reality and are, in particular, responsible for generation of the networks of information systems (data tables) through interactions with the configurations of physical objects. Development of a particular network of information systems is guided by the need to learn the relevant computational building blocks that are necessary for perception, using the formulation by Leslie Valiant. Among these blocks, often learned hierarchically, one can distinguish patterns, clusters or classifiers. The computational building blocks are used by agents, e.g., for approximation of conditions responsible for initiating actions or plans. Agents performing computations based on interaction with the physical environment learn new c-granules, in particular, in the form of interaction rules, representing knowledge not known a priori by agents. These new c-granules are used not only for construction of compound abstract objects but also of compound physical objects, e.g., sensors composed out of more primitive sensors. Learning of interaction rules also supports the control of agents, in particular the self-organized distributed control. Numerous tasks of agents may be classified as control tasks performed by agents aiming at achieving the high quality computational trajectories of configurations of c-granules relative to the considered quality measures over the trajectories. Reasoning supporting agents in searching for solutions of their tasks is based on adaptive judgment, an important component of IGrC. Methods based on adaptive judgment allow agents to construct from given configurations of their c-granules new ones. These new configurations of c-granules should be constructed taking into account the needs of agents realized through interactions with the environment. Here, new challenges are related to developing strategies for predicting and controlling behaviors of agents. We propose to investigate these challenges using the IGrC framework with adaptive judgment used for controlling of computations performed on c-granules. For example, adaptive judgment is used in adaptive learning of rough set based approximations of complex vague concepts evolving with time. It is also used in the risk management of granular computations, carried out by agents, toward achieving the agent needs. XVI A. Skowron