Project "ExSys" was conceived as a creation of expert system for agricultural harvest forecasting. That is why the main attention was paid to the aspect of interaction between man (an expert in the given area) and machine, especially to the interpretability of regularities contained in the data and option of manual entry of prior (expert) information. As the most adequate mean for such interaction, theory of fuzzy sets, which has been used for dozens of years in solving such badly formalized tasks, was selected. Later the field of created complex application was extended and now it can be used for solution of any classification or regression restoration tasks.

Program complex "ExSys" is a tool for designing of fuzzy expert systems in real time. Depending on the presence or absence of expert and on its level of knowledge, complex may work in modes with different degree of human assistance - from manual mode (all system is directly designed by expert) till completely autonomous (system is created automatically during several seconds according to the set of presedents). The use of fuzzy sets theory and methods of fuzzy logic allowed to formulate regularities in terms of reasonable sentences in natural language. The rules are represented as statements in "IF… THEN…" form with logical operators "AND" and "OR". The following rule was generated automatically from the training sample: "IF (there are many wins AND some draws) OR not many dropped balls THEN the rank is high." During recognition the knowledge base is applied to the object and the result is transformed from linguistic terms into numerical form by means of defuzzyfication in order to get concrete point as an answer. There implemented original algorithm of fuzzy rules generation according to the set of precedents (objects with known answers), which allows to discover ALL significant regularities contained in the data. Each regularity is verified by modern methods of mathematical statistics. This helps to avoid catching noise and random fluctuations in training data, which may affect negatively on the quality of recognition and forecast. The developed rule generation algorithm has another good feature. Methods of adaptive correction (such as bagging, boosting etc.) can be applied to it. One of them (AdaBoost) was successfully implemented. This allowed to improve the performance due to the deeper analysis of specific objects which were processed in a wrong way earlier.
   The basic advantages of the system are
º   The option of monitoring the rules discovered by the system and their editing, removing and adding new ones.
º   The option of monitoring and changing the approximate borders and shapes of fuzzy sets.
º   Different defuzzyfication modes allow to solve both classification and regression restoration tasks.
º   Low risk of overfitting. That is why the result got by the recognition of training sample is unbiased estimate of recognition quality of random object taken from the entire set.
º   Creation of knowledge database with simultaneous update in real time during designing the expert system.

  This program complex was successfully tested on a number of real tasks including medical forecast, estimation of housing costs, football teams forecast etc. We constantly modify and update the system in order it to work better and to simplify the interaction with user.