Concepts,methods, and languages for building timely intelligent systems

We describe the ABE/RT toolkit—a set of design, development, and experimentation tools for building time-stressed intelligent systems-and its use for the Lockheed Pilot's Associate application. We use the termtimely systems to refer to systems with hard real-time requirements for interacting with a human operator or other agents with similar time-scales. The ABE/RT methodology is based on a philosophy of rigorous engineering design in which the application developer works to guarantee the system's timeliness by identifying the various events which require timely responses, determining the worst-case frequencies of these events and the deadlines and durations of the tasks that respond to the events, and then verifying that the run-time system has enough processing resources to complete all mandatory taks by their deadlines. We believe this is the only way in the near-term to build complex real-time intelligent systems that will be reliable enough for critical applications with demanding users. The ABE/RT Toolkit contains a set of languages for specifying the structure and behavior of timely systems, together with tools to simulate those models, log and analyze data collected during simulation runs, predict an application's performance on a specified target hardware architecture, and deploy the application on the target architecture.This research was partially funded by the Defense Advanced Research Projects Agency, 1400 Wilson Blvd., Arlington, VA 22209, under contracts F30602-85-C-0135 and F33615-85-C-3804, administered by the Air Force Systems Command, Rome Air Development Center and the Air Force Cockpit Technology Directorate, Wright Research and Development Center, respectively. Use of this material, including copying, by the U.S. government is permitted in accordance with the terms of those contracts.     

ABE: an environment for engineering intelligent systems

The ABE multilevel architecture for developing intelligent systems addresses the key problems of intelligent systems engineering: large-scale applications and the reuse and integration of software components. ABE defines a virtual machine for module-oriented programming and a cooperative operating system that provides access to the capabilities of that virtual machine. On top of the virtual machine, ABE provides a number of system design and development frameworks, which embody such programming metaphors as control flow, blackboards, and dataflow. These frameworks support the construction of capabilities, including knowledge processing tools, which span a range from primitive modules to skeletal systems. Finally, applications can be built on skeletal systems. In addition, ABE supports the importation of existing software, including both conventional and knowledge processing tools     

A domain-specific software architecture for adaptive intelligentsystems

A good software architecture facilitates application system development, promotes achievement of functional requirements, and supports system reconfiguration. We present a domain-specific software architecture (DSSA) that we have developed for a large application domain of adaptive intelligent systems (AISs). The DSSA provides: (a) an AIS reference architecture designed to meet the functional requirements shared by applications in this domain, (b) principles for decomposing expertise into highly reusable components, and (c) an application configuration method for selecting relevant components from a library and automatically configuring instances of those components in an instance of the architecture. The AIS reference architecture incorporates features of layered, pipe and filter, and blackboard style architectures. We describe three studies demonstrating the utility of our architecture in the subdomain of mobile office robots and identify software engineering principles embodied in the architecture     

Ready to leap to neo-creationism

Like the master puppet makers of the classic folk tales, artificial intelligence (AI) engineers have built some marvelous machines. The first 50 years of AI focused on programming computers to perform tasks that previously only humans could do tasks such as speech, driving, planning, problem solving and vacuuming. The principal challenges have been to create suitable language systems and inference engines and then to more or less manually stuff those with sufficient facts, heuristics and algorithms to perform well-defined tasks. To get there, we need a singularity of artificial neo-creationism, where we launch artificial beings into the world that can adapt, learn and evolve themselves. The term neo-creationism means an intentional effort to populate the world with intelligent entities initially engineered by humans. The biblical story of Genesis, the heart of creationism, provides an apt foil for understanding the goals and scope of the R&D program advocated here. To reach the goal reasonably quickly, we should equip these creatures with as much capability and knowledge as possible.     

Puppetry vs. Creationism: Why AI Must Cross the Chasm

Artificial intelligence began with an enthusiastic embrace of newly available computing machinery and the basic question of what kinds of problems we could solve with it. The first 50 years focused on programming computers to perform tasks that previously only humans could do. Then, people began comparing machines to humans as problem solvers, and the race was on to see where machines could match or even surpass human performance. Success in solving math word problems, winning checkers and chess championships, understanding natural language, and generating plans and schedules reinforced our efforts to build supercapable machines. The author call these puppets, not to derogate the machines but to respect the importance of the programmers and builders who were actually responsible for their accomplishments. From time to time, many of us have recognized the field's rate-limiting factor under various names and viewpoints, such as the knowledge-acquisition bottleneck and the challenges of machine learning, system bootstrapping, artificial life, and self-organizing systems. Mostly, however, these efforts have had limited success. The little bit of learning and adaptation they've demonstrated has paled in comparison to the puppeteers' laborious inputs     

Goals and functions of the human body: an MFM model for faultdiagnosis

Describes the use of explicit models of goals and functions for monitoring and diagnosis of intensive-care patients. The method is based on multilevel flow models (MFM) and used in the Guardian system. It provides this system with alarm analysis, fault diagnosis, and automatic generation of explanations. Advantages include a relatively easy knowledge engineering effort and good properties for use in a system with hard real time deadlines. The results of some experiments are also reported     

Architectural foundations for real-time performance in intelligent agents

Intelligent agents perform multiple concurrent taks requiring both knowledge-based reasoning and interaction with dynamic entities in the environment, under real-time constraints. Because an agent's opportunities to perceive, reason about, and act upon the environment typically exceed its computational resources, it must determine which operations to perform and when to perform them so as to achieve its most important objectives in a timely manner. Accordingly, we view the problem of real-time performance as a problem in intelligent real-time control. We propose and define several important control requirements and present an agent architecture that is designed to address those requirements. The proposed architecture is a blackboard architecture, whose key features include: distribution of perception, action, and cognition among parallel processes, limited-capacity I/O buffers with best-first retrieval and worst-first overflow, dynamic control planning, dynamic focus of attention, and a satisficing execution cycle. Together, these features allow an intelligent agent to trade quality for speed of response under dynamic goals, resource limitations, and peformance constraints. We illustrate application of the proposed architecture in the Guardian system for surgical intensive care monitoring and contrast it with alternative agent architectures.This research was supported by DARPA contract N00039-83-C-0136, NIH contract 5P41-RR-00785, EPRI contract RP2614-48, and AFOSR contract F49620-89-C-0103DEF, and by gifts from Rockwell International, Inc. and FMC Corporation, Inc. The Guardian system is being developed in collaboration with Adam Seiver, Rich Washington, David Ash, Rattikorn Hewett, Anne Collinot, Luc Boureau, Angel Vina, Ida Sim, and Michael Falk. The paper's treatment of real-time requirements reflects discussions with colleagues involved in the AFOSR Program on Intelligent Real-Time Problem Solving Systems-especially Stan Rosenschein, Lee Erman, and Yoav Shoham. The paper also benefited from constructive criticism by several anonymous reviewers. Thanks to Ed Feigenbaum for sponsoring the work at the Knowledge Systems Laboratory.