GMDH Application for autonomous mobile robots control system construction A.V. Tyryshkin, A.A. Andrakhanov, A.A. Orlov Tomsk State University of Control Systems and Radioelectronics

Classification of existing autonomous robots

Nearest analog – agricultural AMR Lukas

Basic works on GMDH application to AMR control C.L. Philip Chen, A.D. McAulay Robot Kinematics Learning Computations Using Polynomial Neural Networks, 1991; C.L. Philip Chen, A.D. McAulay Robot Kinematics Computations Using GMDH Learning Strategy, 1991; F. Ahmed, C.L. Philip Chen An Efficient Obstacle Avoidance Scheme in Mobile Robot Path Planning using Polynomial Neural Networks, 1993; C.L. Philip Chen, F. Ahmed Polynomial Neural Networks Based Mobile Robot Path Planning, 1993; A.F. Foka, P.E. Trahanias Predictive Autonomous Robot Navigation, 2002; T. Kobayashi, K. Onji, J. Imae, G. Zhai Nonliner Control for Autonomous Underwater Vehicles Using Group Method of Data Handling, 2007;

Part I Inductive approach to construction of AMR control systems

Problems of AMR design Navigation Obstacle Recognition Autonomous Energy Supply Optimal Final Elements Control Technical State Diagnostics Objectives Execution Knowledge Gathering and Adaptation

Generalized structure of AMR

Objective aspects of AMR control system construction Utility Realizability Appropriateness Classification Taking into account Internal system parameters Forecasting

Features of AMR obstacle recognition Lack of objects a priori information Objects to recognize are complex ill-conditioned systems with fuzzy characteristics Objects are characterized by high amount of difficultly- measurable parameters It is necessary to take into account internal systems parameters for objects classification according to obstacle/not obstacle property, i.e. it isnt possible to find out is this object obstacle or not without regard for system state. There is no necessity to perform full object identification, i.e. it isnt necessary to answer a question What object is this?

Part II Autonomous Cranberry Harvester

Expected Engineering-and-economical Performance Nominal Average AMR speed: Cranberry harvesting coverage: Relative density of harvested cranberry: Total weight of harvested cranberry per season: Season income: $

Automated cranberry harvester

Part III Simulation Results

Object Recognition Data Sample Learning samples – 92; Training samples – 50. Values Ranges: Object Length L Є [0;20] м; Object Width w Є [0;20] м; Object Height h Є [0;20] м;

Recognizing Modified Polynomial Neural Network

Objective Functions Data Sample Learning samples – 140; Training samples – 140. Values Ranges: Surface density of cranberry distribution ρ cranberry Є [0;1] kg/m 2 ; Cranberry harvesting efficiency η Є [20;75] %; Average AMR speed V average Є [0;7] km/h; Nominal average AMR speed V nom average Є [2;4] km/h; AMR engine fuel consumption per 100 km P fuel Є [150;600] liters/100 km. Values laws of variation:

Objective Functions Function of maximal cranberry harvest in preset time: Function of maximal cranberry harvest in minimal time: Function of maximal cranberry harvest with minimal fuel consumption:

Main Indices of Simulation Data CR Percentage of Errors % F(m cranberry,Δt)F(m cranberry,t)F(m cranberry, m fuel ) CRBSCRBSCRBS 3.8e-49.8e-38.6e e ) Obstacle recognition criterion values 2) Objective Functions criterion values

Man should grant a maximal freedom to the computing machinery. Like a horseman having lost a way leave it to a discretion of his horse... A.G. Ivakhnenko. Long-term forecasting and complex system control, Publ. Технiка, Kiev, – p. 8.

Нахождение разделяющих областей в пространстве параметров распознавания

Современные состояние разработок в области АПК

Итерационный алгоритм МГУА с разделением обучения