Options
2001
Report
Titel
Embodied cognitive science, intelligent behavior control, machine learning, soft computing and FPGA Integration. Towards fast, cooperative and adversarial robot team (RoboCup)
Abstract
Roboter, die sich autonom in einer nicht für sie geschaffenen und daher aus ihrer Sicht nichtdeterministischen Umgebung bewegen und bestimmte Aufgaben erfüllen sollen, müssen sich dieser Umgebung anpassen. Solche Umgebungen ändern sich in der Regel permanent, wobei sowohl statische Eigenschaften (z.B. die Position von Gegenständen), wie auch dynamische Eigenschaften (z.B. die Bewegungsweise von Gegenständen) zu berücksichtigen sind. Ein autonomes, lebenslanges Lernen von Robotern kann sich einer Menge von existierenden Techniken bedienen, es bedarf aber auch einer Integration und Anpassung dieser Techniken, um den Anforderungen in der Robotik zu genügen. Dieses Papier stellt zunächst die Problemstellungen und Anforderungen für fußballspielende Roboter vor, die an den Wettkämpfen der RoboCup Federation teilnehmen. Ebenso werden die bestehenden Arbeiten in der RoboCup-Gruppe des Institutes AiS beschrieben. Anschließend wird die Notwendigkeit des lebenslangen Lernens und damit verbundene Fragen grundsätzlich behandelt. Der Hauptteil dieses Papiers beschreibt wie mit Hilfe von künstlichen neuronalen Netzen (NN) die Erfolgsaussichten von Verhaltensweisen des Roboters für unterschiedliche Situationen gelernt werden können. Darauf basierend kann ein Roboter in einer bestimmten Situation das Verhalten auswählen, das die höchsten Erfolgsaussichten verspricht. Für die RoboCup-Roboter von AiS wurden exemplarisch für eine Situation (Roboter ist im Ballbesitz und fährt Richtung generisches Tor) die Erfolgsaussichten gelernt (off-line). Dieses gelernte Wissen wurde in die Verhaltensprogramme der AiS-Roboter integriert und während der RoboCup Weltmeisterschaften 2000 in Melbourne, Australien erfolgreich eingesetzt. Um das Verhalten eines Roboters in möglichst vielen Aspekten an die Umwelt anzupassen und dies möglichst während der Ausführung des Verhaltensprogrammes (on-line), muß das Lernen hoch performant und in Real-Zeit durchgeführt werden. Daher wird in dieser Arbeit das Lernen mit künstlichen neuronalen Netzen (NN) in eine FPGA Architectur integriert. Abschließend werden in diesem Papier alternative Ansätze und weitere Entwicklungsmöglichkeiten des von uns gewählten Ansatzes diskutiert.
;
The interest in complete systems (robots), i.e., autonomous, self-sufficient, embodied, and situated rather than only on a specific part (physical or theoretical) is becoming necessary from embodied cognitive science and understanding natural intelligence perspectives. Robots have also to be reactive when generating actions (behaviors), i.e., at a rate that is commensurate with the dynamics of environments in which they are embedded. This has motivated, in robotics and particularly in robotic soccer, several researchers to underline the necessity of behavior learning, lifelong learning, opening then the opportunity to learned knowledge transfer. By another way, soft computing, involving Intelligent Systems (IS) and Hybrid Intelligent Systems (HIS) implying particularly Neural Networks (NN), Fuzzy Logic (FL), Genetic Algorithms (GA), and Adaptive Resonance Theory (ART) has been recognized to improve the learning, adaptation, generalization, and prediction capabilities from past experiences in real dynamics environments. The aim of this work focalizes on the intelligent behavior control leading to intelligent behaviors achieving a fast, cooperative and adversarial robot team (RoboCup) in order to provide GMD mobile robots with more autonomy, intelligence, and real-time processing capabilities. First, an overview is given on both Robot World Cup Initiative (RoboCup) and RoboCup project developed at AiS-GMD. Second, the problem definition and objectives, of this work, from embodied cognitive science and understanding natural intelligence perspectives are presented. Third, NN based learning of elementary behaviors and their integration in FPGA architectures for a fast moving robot team (RoboCup) is suggested. Indeed, just as young soccer players must learn to control the ball before learning any complex strategies, robots must also acquire low-level skills before exhibiting complex behaviors: the most sophisticated understanding of how to act as part of a team is useless without the ability to execute the necessary individual behaviors. Thus, to react in real-time, FPGA architectures characterized by their high flexibility and compactness are suggested for the NN integration. Afterwards, behavior learning to predict using NN: towards a fast, cooperative and adversarial robot team (RoboCup) is suggested to enhance the elementary behavior "Kick". In fact, for instance, through experience one might learn to predict for particular game situations (e.g., chess positions) whether they will lead to a win. In robotic soccer, one might also learn to predict for particular situations (e.g., teammate and opponent positions) whether they will lead to a success. Thus, the suggested NN-prediction demonstrated, during the 4th World Championships RoboCup'2000, cooperative and adversarial behaviors especially face to situations where the successfulness of "Kick" is not guaranteed. With such intelligent elementary behavior "Kick", GMD mobile robots involve multiple players, implying then cooperation and coordination among the players of the team (teammates) and taking into account the players of the adversary team (opponents). By another way, the signals of sensors are often noisy or they are defective giving incorrect data. This problem is efficiently handled by NN with their inherent features of adaptivity and high fault and noise tolerance making them robust. Thus, FPGA architectures for elementary behaviors with their high operating speeds, low power consumption and reprogrammable features as well as NN-Prediction provides GMD mobile robots with real-time processing capabilities, more autonomy, intelligence, and flexibility making them more robust and reliable and consequently lead them towards a fast, cooperative and adversarial robot team (RoboCup). Finally, in order to achieve fast, cooperative and adversarial robot team (RoboCup), some interesting alternatives for future works are presented related to 1) soft computing, involving IS and HIS, based learning of elementary behaviors and their integration in FPGA architectures for fast moving robot team ; 2) behavior learning to predict using soft computing involving IS and HIS: towards fast, cooperative and adversarial robot team ; 3) Dual Dynamics Simulator (DDSim) as pedagogical tool for soft computing involving IS and HIS implying particularly NN, FL, GA, and ART, e.g., NN, Fuzzy Neural Networks (FNN), and Fuzzy ArtMap Neural Networks (FAMNN) ; and 4) reinforcement learning (Q-Learning) and lifelong learning (Learned Knowledge Transfer).