1、FROM INTELLIGENT CONTROL TO COGNITIVE CONTROLKAZAHIKO KAWAMURA, CENTER FOR INTELLIGENT SYSTEMS, VANDERBILTUNIVERSITY, USA, KAZ.KAWAMURAVANDERBILT.EDUSTEPHEN GORDON, CENTER FOR INTELLIGENT SYSTEMS, VANDERBILTUNIVERSITY, USA, STEPHEN.M.GORDONVANDERBILT.EDUABSTRACT This paper describes our efforts to d
2、evelop a robot with robust sensorimotor intelligence using amultiagent-based robot control architecture and a biologically inspired intelligent control. Such control is called cognitive control. In this paper we will discuss the application of cognitive control to a humanoid robot. Features of cogni
3、tive control addressed include short-term memory for environmental learning, long-term memory for behavior learning and task execution using working memory and TD learning.KEYWORDS: intelligent control；cognitive control；behavior learning； working memory1. INTRODUCTION IEEE Control Systems Societys T
4、echnical Committee on Intelligent Control states that The area of intelligent control is a fusion of a number of research areas in systems and control,computer science, and operation research among others, coming together, merging and expanding in new directions . 1. A key sentence here is fusion of
5、 research areas. As the need to control complex systems increases, it is important to look beyond engineering and computer science areas to see if we can learn from them. For example, humans have the capacity to receive and process enormous amount of sensory information from the environment, exhibit
6、ing integratedsensorimotor intelligence as early as two years old 2. A good example of such sensorimotor intelligence by adults is the well-known Stroop test 3, Appendix 9.1. Thus it is a challenge for control engineers to find ways to realize humans robust sensorimotor mechanisms called cognitive c
7、ontrol 4 within machines. Most goal-oriented robots currently perform only those or similar tasks they were programmed for and very little emerging behaviors are exhibited. What is needed is an alternative paradigm for behavior learning and task execution. Specifically, we see cognitive flexibility
8、and adaptability in the brain as desirable design goals for the next generation of intelligent robots. Several cognitive architectures have been implemented for the purpose of testing human psychological models5 6, but such models have not been fully adopted by the robotic community.2. INFORMATION P
9、ROCESSING IN HUMANS Engineers have long used control systems utilizing feedback loops to control mechanical systems. Figure 1 illustrates a class of adaptive (or learning) control systems 7. Limitations of model-based control led to a generation of intelligent control techniques such as fuzzy contro
10、l,neuro computing and reconfigurable control 1. The human brain is known to process a variety of stimuli in parallel, ignore non-critical stimuli to execute the task in hand, and learn new tasks with minimum assistance. This process,known as executive or cognitive control, is unique to humans and a
11、handful of animals 2. Figure 2 illustrates a conceptual model of cognitive control in which we are using to realize robust behavior generation and learning for our humanoid robot.Figure 1. An adaptive control system 8Figure 2. Model of cognitive controlModified from Miller, et. Ala93. MULTIAGENT-BAS
12、ED COGNITIVE ROBOT ARCHITECTURE As the complexity of a task grows, so do the software complexities necessary to process sensory information and to control actions purposefully. Development and maintenance of complex or large-scale software systems can benefit from domain-specific guidelines that pro
13、mote code reuse and integration through software agents.Information processing in our humanoid robot ISAC (Intelligent Soft Arm Control)is integrated into a multiagent-based software architecture based on the Intelligent Machine Architecture (IMA) 9. IMA is designed to provide guidelines for modular
14、 design and allows for the development of subsystems from perception modeling to behavior control through the collections of IMA agents and associated memories, as shown in Figure 3 10.ISACS IMA-based Cognitive ArchitectureFigure 3. Multiagent-based cognitive robot architecture ISACS memory structur
15、e is divided into three classes: Short-term memory (STM), long-term memory (LTM),and the working memory system (WMS). STM holds sensory information of the current environment. LTM holds learned behaviors, semantic knowledge, and past experiences.WMS holds task-specific information called chunks and
16、streamlines the information flow to the cognitive processes during the task execution. STM is implemented using a sparse sensory data structure called the Sensory EgoSphere (SES). SES, inspired by the egosphere concept defined by Albus 11 I, serves as a spatio-temporal STM for a robot 12. LTM stores information such as skills learned and experiences gained for future retrieval.4. COGNITIVE C
