Artificial Intelligence System Designer 4 GN ISP RAS

Artificial Intelligence System Designer 4 GN ISP RAS Alexander Zhdanov

Artificial Intelligence Problems solved by means of AI systems: ® Pattern recognition ® Data mining ® Image recognition ® Automated reasoning ® Expert systems ® Prediction ® Automated control Approaches used: • Artificial Neural networks • Fuzzy logic • Reinforcement learning • Stochastic approaches • Structural representation

AAC Framework for AI system design Explicitly deals with Well-suited for: ® Pattern recognition ·Pattern recognition ® Knowledge base formation systems ® Prediction ·Expert systems ® Automated analysis ·Data mining ® May be based on system ® Determined chaos systems ·Adaptive control ® 3 rd generation neural networks systems ® Genetic algorithms ® ® ® Stochastic methods Open for other sophisticated techniques

AAC Framework for AI system design AAC Comparison with other AI-approaches: Artificial neural networks - perform only patter recognition or approximation and demand a priori learning. AAC systems have abilities for self-control ® Fuzzy logic systems – demand a priori formulated fuzzy rules. AAC systems deduce rules themselves and corrects them if necessary ®

AAC Framework for AI system design For example, the Ad. CAS system for car suspension adaptive control could not be created simply on basis of another method: artificial NN, reinforcement learning, fuzzy logic or any another approach.

AAC Framework applicability ® Pattern recognition systems ® Prediction, forecasting systems ® Expert systems, decision making ® Adaptive control systems ® Highly adjustable to problem domain or context

Universality of a system “Applicability of a method is inversely proportional to its universality” It is impossible to create universal control system for ANY customers and ANY problem, because its parameters have to depend on given objects • Parameters of CS, which are independent from CO 1) The structure of the CS operation; 2) The ways in which the CS subsystems are constructed – the recognition system, knowledge base and other subsystems; 3) The models of the neuron-like elements of which the CS subsystems are constructed; etc. • Parameters of CS, which dependent from CO 1) The input and output variables and their characteristics; 2) The rules of pattern formation which will be required for the control; 3) The rules of knowledge formation in the knowledge base 4) The qualitative criteria for the evaluation of the possible states of the CS, for the control quality evaluation, for the determination of the goal functions; etc.

Computer Aided System Engineering ® Pros ® Produces highly customizable solutions ® Ease of use: does not require hardcore programming skills ® Adaptability ® Flexibility ® Development of end-user solutions ® Cons (requires from developers) ® High level of abstraction on analysis stage ® Deep understanding of system principia ® Ability to translate abstraction into concrete notions

CASE for design of AI systems Computer model of given application Model of sensors AAC method Programming Prototype of CS Model of CO Model of actuators The software Tools

Proposal #1: Development of CASE for design of AI systems based on AAC method Main features: ± ± ± ± Drastically reduces time and resources required for development Explicit AI orientation Export/import interfaces with simulation software and with hardware Advanced visualization and analysis techniques Easy to use for non-experienced programmer Makes process of AI system design more transferable Open interfaces

Three main phases of applied AI systems projecting for various tasks Control System Phase I. CS Controlled Object Sensors CO Actuators Environment l The projection and development of the CS prototype and its testing on program models of the CO, Sensors and Actuators. (1 -2 years) Phase II. CS Sensors CO Actuators Environment l The debugging of the CS prototype on real CO, sensors, actuators or their physical models. Phase III. CS Sensors CO Actuators Environment l Building-in of the CS into the real CO, where the CS is implemented on real (on-board) processor.

Proposal #2: AAC-based multiagent system architecture Development of multiagent system architecture based on AAC Main features of such system: • • Explicitly distributed AI-based (in sense of AAC) Self-monitoring, Self-adaptability, Self-manageability, Selflearning Secure Adaptable to heterogeneous environment Decentralized control Flexibility Eco-system principles based

Thank you for your time Welcome to discussion