of Cognitive Modeling MIPT
In the scientific direction of artificial intelligence, one of the priority tasks is the task of creating universal control systems for intelligent agents (software and robotic). By universality, we mean, firstly, the integration of different methods for solving one or several intellectual problems, and secondly, providing an agent with the ability to operate in a wide range of complex dynamic environments. Recently, the development of artificial intelligence is moving towards the specialization of technologies, their adaptation to solve narrower and simpler problems. However, recently, in the direction of general artificial intelligence (AGI), there has been a revival of the original ideas aimed at creating universal intelligent systems capable of solving a wide range of problems.
The Center for Cognitive Modeling develops one of such universal approaches - cognitive computer modeling. We propose to combine such methods as behavior planning, knowledge representation, reasoning modeling, reinforcement learning and computer vision based on deep neural network architectures to build a universal cognitive agent control system that can operate in both virtual and physical environments, solving assisting problems. a person in a wide range of conditions (home, office, industrial environments).
In our approach, we propose to decompose the agent management system into three levels: reactive, tactical and strategic - and develop the STRL (strategic tactic reactive layered) architecture. Accordingly, in the Center for Cognitive Modeling, groups are distinguished according to various subtasks solved at these levels.
At the reactive level, the problems of the theory of controlling a model of a physical object (real or virtual) are solved; in fact, this is the integration of the equations of dynamics. Subproblems of stabilization, trajectory following, change of control modes, etc. are solved here. Many of these subtasks have quality efficient implementations that come with a specific robotic platform.
At the tactical level, the tasks of computer vision, localization, mapping and navigation in space are solved. At this level, local specialized subsystems are trained, including those implemented on the basis of deep neural network architectures, for segmentation and classification of images, building a map of the area, highlighting objects of interest, etc.
The strategic level contains subsystems for representing conceptual knowledge, which are associated with low-level tactical representations, a scheduler, and a learning subsystem, including those based on reinforcement learning.
The main purpose of the Center's work is both the creation of a theoretical basis for the construction and operation of such behavioral control systems by a cognitive agent, and the creation of software systems for solving applied problems. We believe that interdisciplinarity and general theoretical methods allow for breakthroughs in fundamental and applied work on artificial intelligence. This, in turn, will help us better understand how our own natural intelligence works.
Our Center is constantly developing, launching new applied and academic projects. We are always glad to motivated, ideological guys who want to prove themselves in one of the most promising areas of modern science.