Letters to the Editor

The Phoenix project (“Trial by Fire: Understanding the Design Requirements for Agents in Complex Environments.” AI Magazine, Vol. 10, No. 3) presents very interesting work in forest fire simulation. I am especially glad to see recognition that the “realtime, spatially distributed, multiagent, dynamic, and unpredictable fire environment” provides an excellent opportunity to explore a variety of AI issues, such as how complex environments constrain the design of intelligent agents. I hope more AI researchers will venture into the complex domain of forest fire management. However, I am concerned about the fire knowledge presented in the article. In defending their use of simulation, Cohen et al. argue that their simulated fire environment is a complex environment, irrespective of whether it is an accurate model of how forest fires spread. Yet, they recognize the importance of a realistic simulation, “the point of using simulators is to create realistic and challenging worlds.” Then they make the unsubstantiated claim that the “fire environment is an accurate model of forest fires.” Mathematical models of fire spread have been developed that do a reasonably good job of prediction, if the models are applied within the bounds of the simplifying assumptions. That is a very big if. To state unequivocally that an accurate model of fire spread exists is somewhat of an overstatement. Examining the fire simulation shown in figures 1 through 4, the fire appears to spread from the northwest to the southeast in a somewhat elliptical shape. This appears reasonable and is consistent with other models of fire spread. However, this would mean that the rear of the fire is in the northwest corner. The starting position of the bulldozers indicates that the rear of the fire is in the northeast corner. Something is amiss. In addition, the length to width ratio of the fire is about 1.5, indicating a mid-