This paper will detail the progress on the development of the Symbolic and Subsymbolic Robotics Intelligence Control System (SS-RICS). The system is a goal oriented production system, based loosely on the cognitive architecture, the Adaptive Control of Thought-Rational (ACT-R) some additions and changes. We have found that in order to simulate complex cognition on a robot, many aspects of cognition (long term memory (LTM), perception) needed to be in place before any generalized intelligent behavior can be produced. In working with ACT-R, we found that it was a good instantiation of working memory, but that we needed to add other aspects of cognition including LTM and perception to have a complete cognitive system. Our progress to date will be noted and the challenges that remain will be addressed.
Keywords: cognitive architectures, robotics, cognition, working memory, context
1. INTRODUCTION
SS-RICS was developed by the U.S. Army Research Laboratory’s (ARL) Human Research and Engineering Directorate (HRED) in cooperation with Towson State University beginning in 2004. The goal of the program was to develop a system capable of performing a wide variety of autonomous behaviors under a variety of battlefield conditions.
As a general theoretical position, we have taken the stance that cognition arises from a collection of different algorithms, each with different functionalities, which together, produce the integrated process of cognition. This is also known as a functionalist representation [1]. We are developing SS-RICS to be a modular system, or as a collection of modular algorithms, each group of algorithms with different responsibilities for the functioning of the overall system. The important component is the interaction or interplay amongst these different algorithms, which leads to an integrated cognitive system. We are not necessarily attempting to produce a neurological representation of the individual components of the brain (thalamus, amygdale), but instead, a functional representation of cognition (learning, memory).
We began the development of SS-RICS by using the existing cognitive architecture ACT-R [2] as a framework. ACT-R has a long history of development and has continued to be refined to the present day. ACT-R grew from the artificial intelligence (AI) symbolic tradition beginning with Newell and Simon [3] and their work on the generalized problem solver (GPS) and their studies of the problem solving strategies of chess masters. ACT-R grew from the production system based work of GPS and was later augmented to include memory algorithms developed by John Anderson, and subsequently to include a variety of additional learning methods and algorithms. ACT-R has been used primarily to simulate human performance data and to make predictions of human performance and error data [4]. ACT-R currently enjoys a large user base and undergoes continuous revisions and improvement.
In our work on the development of SS-RICS, we found that ACT-R was a sufficient approximation of working memory for a robotic system, but that the robot needed other systems in order to function in a dynamic world, primarily perceptual systems and LTM. The perceptual systems within ACT-R were not sufficient for robotic control and the LTM mechanisms were model specific and not generalized knowledge. Consequently, we added LTM to SS-RICS by using ConceptNet [5] and we added a variety of perceptual systems to SS-RICS by using statistical techniques and neural networks.
DOWNLOAD...
Troy Dale Kelley and Eric Avery
U.S. Army Research Laboratory, Aberdeen Proving Ground, Aberdeen, MD 21005
troy.kelley@us.army.mil eric.s.avery@us.army.mil
No comments:
Post a Comment