Annual Review in Automatic Programming最新文献

An architecture for execution supervision of Robotic Assembly Tasks is presented. This architecture provides, at different levels of abstraction, functions for dispatching actions, monitoring their execution, and diagnosing and recovering from failures. Modeling execution failures through taxonomies and causal networks plays a central role in diagnosis and recovery. A discussion on the knowledge acquisition process, through the use of machine learning techniques, is made. Preliminary results in this area are presented and planned extensions discussed.

Adaptive intelligent agents pursue multiple goals in dynamic, complex, uncertain environments. Illustrative applications include medical monitoring agents, autonomous robots, and animated actors. Real-time considerations include hard and soft deadlines, as well as more qualitative constraints. There are no bounded-time algorithms for such applications. Agents must construct goal-seeking behavior opportunistically out of component behaviors triggered at runtime. In contrast to conventional real-time systems, an agent's real-time effectiveness cannot be guaranteed by a single provably optimal technique. Instead, it must emerge from interactions among multiple heuristic techniques at architectur, cognition, and control levels.

To exploit the potential offered by integrated intelligent control systems, it is essential that underlying architectures are available that are able to integrate not only hardware and software, but also human beings into the control loop. By studying the way that a human workforce operates, this paper suggests how a human systems analogy can be drawn up that highlighted some fundamental differences between the way that humanbased, and computer-based systems operate. One of the key differences is that humans are able to reason not only logically, but also in terms of time. This paper addresses the philosphy behind the DENIS architecture — a distributed architecture that enables this high level of compatability between autonomous intelligent agents.

The following paper describes efforts to develop a processor architecture that meets the requirements of hard real time computing. The architecture is of the RISC-type with a single, modular CPU. The modules are a Kernel Processor, a Task Processor, a Memory Module and a Controller for internal and external communication. By integrating multiple register files directly accessible by the ALU, the number of main memory accesses decreases and the time for context-switches is reduced considerably. While OS functions, scheduling, time management and interrupt handling are performed by the Kernel Processor, the Task Processor focuses on its primary function, viz., to execute application program code. Assigning the traditionally sequentially performed program-, operating system- and memory-operations to different modules working in parallel results in a significant increase of performance. The reduced instruction set interfacing this architecture allows for a complete and convenient implementation of real time algorithms, especially in distributed systems, without loosing the operational determinism, which was one of the major design guidelines.

Continuous state plants place specific demands on the structure and operation of multi-agent, multi-paradigm distributed intelligent controllers. A controller for a deep shaft mine winder is proposed. Its design demonstrates the choice of agents and how they interact to control a continuous-state plant. The controller is evaluated by means of a simulation study. The study shows that it is helpful to make a distinction between a priori and operational knowledge in designing intelligent controllers for continuous-state plants.

This paper introduces the REAKT (REA1 time Knowledge Tool) tool-kit from the implementation point of view. The Tool-kit is based on a blackboard model of multi-agent cooperation. The paper will thus deepen into some of its architectural characteristics, leading emphasis on practical issues. A particular C++ REAKT Model of Process Concurrency which makes extensive use of object-oriented techniques such as inheritance and polymorphism will be presented. The distribution of application features over independent processes, and the problem of scheduling periodic, and aperiodic processes in a REAKT-based real-time application will be also introduced. Finally, some figures about the performance of key toolkit features will be provided.

On-line process fault diagnosis using fuzzy neural networks is described in this paper. The fuzzy neural network is obtained by adding a fuzzification layer to a conventional feed forward neural network. The fuzzification layer converts increments in on-line measurements and controller outputs into three fuzzy sets: “increase”, “steady”, and “decrease”. Abnormalities in a process are represented by qualitative increments in on-line measurements and controller outputs. These are classified into various categories by the network. By representing abnormalities in qualitative form, training data can be condensed. The fuzzy approach ensures smooth transitions from one fuzzy sets to another and, hence, robustness to measurement noise is enhanced. The technique has been successfully applied to a CSTR system.

The logical design of a neural controller is achieved by representing a neural computation as a stochastic timed linear proof with a built-in system for rewards and punishments based on the timeliness of a computation performed by a neural controller. Logical designs are represented with stochastic forms of proofnets and proofboxes. Sample applications of the logical design methodology to the truck-backer upper and a Real-Time object recognition and tracking system (RTorts) are presented. Performance results of the implementation of the target dynamics identification module of the RTorts are given and compared to similar systems.

A neurofuzzy system combines the positive attributes of a neural network and a fuzzy system by providing a transparent framework for representing linguistic rules with well defined modelling and learning characteristics. Unfortunately, their application is limited to problems involving a small number of input variables by the curse of dimensionality where the the size of the rule base and the training set increase as an exponential function of the input dimension. The curse can be alleviated by a number of approaches but one which has recently received much attention is the exploitation of redundancy. Many functions can be adequately approximated by an additive model whose output is a sum over several smaller dimensional subrnodels. This technique is called global partitioning and the aim of an algorithm designed to construct the approximation is to automatically determine the number of submodels and the subset of input variables for each submodel. The construction algorithm is an iterative process where each iteration must identify a set of candidate refinements and evaluate the associated candidate models. This leads naturally to the problem of how to train the candidate models and the approach taken depends on whether they contain one or multiple submodels.

This paper studies the uncertainty in mobile robot navigation. The paper presents a model to propagate the uncertanty in position and orientation when tracking a given path. The model assumes normal distribution with zero means and a covariance matrix which can be computed recursively using the kinematics. The paper also presents the application to the mobile robot RAM-1 designed and built for navigation in outdoor and indoor industrial environments. The proposed method can be used in the navigation system to know the position uncertainty before the vehicles executes a given path. Furthermore, the method is also useful to plan the execution of computationally intensive vision and other external perception functions which are required to avoid the uncertainty growing during navigation.