![]() |
|||
![]()
|
![]() |
![]() |
![]() |
3.2. PHOENIX: A REAL-TIME, ADAPTIVE PLANNER THAT MANAGES FOREST FIRES IN A SIMULATED ENVIRONMENTCohen et al. PHOENIX represents the traditional line of expert systems in the sense that it came out of university labs as a typical R&D system. But like its academic predecessors, it broke new ground, both as a planning system and as a forest management system. It departs from the major line, first of all for being a simulator that advises by showing and demonstrating rather than concluding and explaining. The PHOENIX project (Cohen 89) was directed toward three main goals, each addressing ground-breaking aspects of both AI planning, knowledge engineering, and system architecture. Its scope included:
Given the last goal, PHOENIX is a tribute from the R&D community to those struggling with the issues of building intelligent systems for practical, real-world operations. PHOENIX was an investigation into management problems typically related to that of administering natural resources. Its task was to control simulated forest fires by deploying simulated bulldozers, crews, airplanes, and other objects. The target area was the Yellowstone National Park. Fires can spread in irregular manners, in various shapes, and at variable rates. This is governed by the state of the environment in which the blaze takes place. Ground cover and elevations, moisture content, wind speed and direction, as well as natural boundaries, are among the most important factors determining the behavior of forest fires. According to Cohen et al., PHOENIX simulates forest fires with accuracy. The system is also capable of simulating firefighting objects such as people and bulldozers in a close to real-life manner. Forest fires are fought by removing one or more of the elements that keep them burning. This can be fuel, heat, and/or air. Cutting a fire line through the forest removes fuel. Water dropped on the fire removes heat. Air can be expelled by means of a fire retardant. Fires of some scale are fought by setting off backfires that are meant to intercept wild fires and denying them fuel. All of this has to be coordinated. For major fires, there can be several geographically dispersed fire bosses and hundreds of firefighters. PHOENIX is not capable of handling a full-scale fire. It restricts itself to fires that can be handled with one fire boss and a few firefighting objects. But its capabilities include abilities to exploit natural firebreaks such as rivers, to initiate backfires, to apply common firefighting plans using bulldozers to isolate fires, and to adjust to new and changing information about weather and fire development. PHOENIX was built using a set of specialized, semiautonomous agents such as bulldozers that carry a distinct purpose, its own knowledge base, an acting script, and protocols for communication and response. Agent design in PHOENIX was circumscribed by what the authors called the "Behavioral Ecology Triangle." The triangle addresses the relationships between environmental characteristics and necessary agent behavior. The connections are obviously given by the design of the agent. As such, PHOENIX touches upon some fundamental epistemologic of knowledge-based design systems. In terms of the forest fire, the agents are subjected to a dynamic, ongoing, and changing world that is both unpredictable and varied. This can be illustrated with changing winds, humidity, and fuel type. It represents a continuous flow of problems, not merely a single challenge (like many expert systems have traditionally focused on) that can be dealt with once and then abandoned. The fire will pose the problems in real-time and in an ever-changing manner that can be truly unexpected. Agents have to cope with multiple scales of time and spatial distribution. The agents involved must be able to prognosticate actions over time and make attempts to foresee developments. Bulldozers do not have unlimited amounts of gasoline and will thus have to determine actions in accordance with the available resources. Uncertainty is a major factor for all agents and can be illustrated by the decision situations to be faced. It is impossible to gain a perfect overview. Information can be scarce. All this requires acute resource management, control of uncertainty, cooperative capabilities, and the ability to plan. PHOENIX is part of an experimental bench that has four layers. It consists of a coordinator, the environment, the agents, and the organization of agents. The coordinator takes care of the actual simulation, maintaining the temporal records of fire development and the movements of the different agents involved. The environment layer handles the factual representation of the fire settings, including maps of Yellowstone National Park. The agents layer contains the agent design. Compared to a traditional expert system using rules, these agents can be conceived as the embodiment of firefighting knowledge. But in addition to specialized domain knowledge, each agent is equipped with a meta-knowledge protocol that evaluates its own problem-solving strategy in terms of overall firefighting objectives. This protocol also includes a communication capability, an ability to replace goals issued by superior agents, and knowledge to coordinate its own actions with others. Like SPECTRUM, agents carry stored, skeletal plans that are stored in a library. When an agent is faced with a particular objective, a matching plan is retrieved, instantiated, and placed on the timeline. The organizing layer defines the operating structure among agents. Agents are specialized, but a force of agents can be organized in different ways and along a scale of different authority and capability. An organizing layer represents a particular knowledge profile capable of doing something better than can be achieved with other configurations. The fire boss can be made the principal agent for many configurations. In the most centralized version, no lateral communication is allowed; everything goes through the boss. The system was built using a Texas Explorer Lisp Machine. Implementation employed object-oriented techniques.
|
![]() |
|
Use of this site is subject certain Terms & Conditions. Copyright (c) 1996-1999 EarthWeb, Inc.. All rights reserved. Reproduction in whole or in part in any form or medium without express written permission of EarthWeb is prohibited. Please read our privacy policy for details. |