1 Algorithms of time compression and analysis of formed patterns in autonomous adaptive control systems Mazur Yuri, Zhdanov Alexander Lebedev Institute of Precision Mechanics and Computer Engineering, Moscow, Russia Autonomous Adaptive Control Lab (AAC Lab)

2 Structure and functions of «nervous system» Autonomous Adaptive Control (AAC) Recognition and pattern formation subsystem Knowledge base Decision making subsystem Emotions apparatus Subsystem of knowledge base generation Memory of recognized patterns Sensors Actuators Environment Body of controlled Object Nervous system (only main elements shown)

3 Memory module for recognized patterns Adding patterns to the memory Fixed size of recognized patterns` memory TNTN Trash T N+1 T N+2 T N+4 T N+3 Deleting patterns from the memory

4 Problems of recognized patterns` module and solutions Problems: Unreasonable using of memory (saving of all input data). Absence of information's compression. Solutions: Compression of information in the time direction. Introduction of importance characteristic of every pattern for control system (CS) and using of this characteristic for compression. Reason: Storing of maximally important information for CS helps to increase further analysis's efficiency of this information and deduction of new regularities, which, probably, accelerates learning process and increase control quality.

5 Time-based compression of information T Initial patterns sequence Needs 9 memory cells Needs 4 memory cells T Compressed patterns sequence TNTN T N+8 T N+4 T N+6 T N+5 T N+3 T N+2 T N+1 T N+7 Compression of data on the basis of average values Adding patterns to the memory Adding patterns to the memory

6 Consideration of information importance for control system in compression process In the AAC method role of information importance (pattern/event) is played by emotional estimation of pattern, because function depending from emotional estimation is used as criterion of control quality.

7 Using of emotional estimation for compression of information Level for all patterns Level 2 for patterns with high emotions Level 0 for patterns with low emotions Level 1 for patterns with medium emotions Emotions intervals all 1 |S|

8 Features of the developed algorithm Averaging function is used for attributes values of compressed patterns. Updating of compressed patterns is happened on each time point. Fixed and preliminary determined memory size is used for intermediate calculations. Level of compression depends on pattern importance for control system and amount of time passed from the moment of pattern recognition.

9 Computer implementation of the algorithm Level for all patterns Level 0 Level 1 Input sequence Emotions

10 Thank you for your attention! Any questions?