MARINA DEL REY, Calif. -- On a quiet street in a village in the Balkans, an accident suddenly puts an American peacekeeping force to the test. A Humvee has hit a car, and a child who has been injured in the collision lies unmoving on the ground. A medic leans over him. The child's mother cries out. A crowd of local residents gathers in the background. How they will react is anyone's guess.

A lieutenant arrives at the scene and is confronted by a number of variables. In addition to the chaos unfolding in the village, a nearby unit is radioing for help. Emotions — not only the lieutenant's own and those of his sergeant, but also those of the panicked mother and the restive townspeople — will clearly play a role in any decision he makes.

This seven-minute situation is a simulation, generated on a large computer screen with sophisticated animation, voice synthesis and voice recognition technology. It is the product of about six months of work here by three research groups at the University of Southern California: the Institute for Creative Technologies, largely financed by the Army to promote collaboration among the military, Hollywood and computer researchers; the Information Sciences Institute; and the Integrated Media Systems Center.

The only human player is the lieutenant. The rest of the characters, including the sergeant who has been conferring with the lieutenant, have been generated by the computer.

For years, the Pentagon has used computers to simulate a wide range of battlefield experiences as realistically as possible. The Army has simulators that allow hundreds of soldiers in different locations to engage in a tank battle on the same virtual battlefield. And the more sophisticated the models, the more adeptly they take into account the effects of noise, vibration, heat and long hours of marching on an individual or group.

But the modeling has gone only so far, confined for the most part to what the Pentagon calls doctrine, or prescribed procedure. A computer-generated mission rehearsal that models human behavior and the emotions that govern it crosses a military frontier.

"It's a complicated, unexpected situation where you have to deal with individuals and not just tanks," said Dr. Paul S. Rosenbloom, a computer scientist who is deputy director of the Intelligence Systems Division at the Information Sciences Institute.

Dr. Richard W. Pew, a principal scientist at BBN Technologies in Cambridge, Mass., and an expert on military simulations, said, "It's a big enough challenge to model the doctrine, but ultimately we'd like to model the variations because the reason these models are potentially useful for training is they behave the way real people do."

Such simulations are still experimental. But when they are ready, they will be used at bases around the country to train soldiers and officers alike to make decisions under stress.

The University of Southern California exercise illustrates the latest challenge among researchers: to focus on the more unpredictable side of the human psyche, simulating emotions and the unexpected effects that panic, stress, anxiety and fear can have on actions and decisions when an officer or a soldier is deep in the fog of war.

"How does a hand-grenade explosion a few feet away from you motivate you if you've just been marching 16 hours in tremendous heat?" asked Dr. Barry G. Silverman, an engineering professor at the University of Pennsylvania.

Of course, video and computer games are the closest most people come to experiencing situations like that. In fact, Dr. Silverman said one of his students had recently asked him why he even bothered with his research when there are games like Age of Empires, Microsoft's popular warfare strategy series.

The similarities, Dr. Rosenbloom acknowledged, are obvious. "You're in an environment and interacting with characters trying to accomplish something," he said. "The difference is in the underlying technology."

But there are also important distinctions. Simulations like those he works on, he said, are more concerned with trying to understand how people really react in a variety of circumstances.

"It can react appropriately to you under a wide range of things you might do and essentially intelligently adapt its behavior," he said. "There are no computer games that have anywhere near that flexibility. There aren't characters who have a deep understanding of a situation."

Dr. Silverman agreed. "Most of the games out there are artistically and stylistically impressive, but not entirely faithful to real human behavior," he said. "We can't take those games and easily replace their made-up forces with ones we'd like to fight against."

Current models of human behavior are known to be brittle, in computer simulation parlance, which means they are unreliable. "They can't react to unanticipated circumstances," said Dr. Harold L. Hawkins, who oversees behavior modeling programs for the Office of Naval Research, "and they lack the adaptability and basic cognitive and perceptual functions all of us take for granted. The goal is to try to improve the realism to minimize the brittleness."

The growing interest among researchers in these kinds of simulations comes with the rise in computer processing power and the growing sophistication of psychological theories. Dr. Rosenbloom said that tasks that might have required special computers five years ago could now be done with ordinary machines. "It doesn't matter what kind of computers they are," he said. "It just matters that they're getting faster."

Dr. Silverman said that not only have new theories emerged on the role of emotions in decision-making, but computer processing has also become fast and inexpensive enough that computers are able to do the complex calculations necessary to model behavior. "The agents can now think in real time and react and have emotions in real time," he said.

But the degree to which emotions can be simulated depends on the type of behavior being modeled. The effect of stress or sleep deprivation, for instance, is better understood than, say, panic, fear or anxiety.

Wayne Zachary, chief executive of Chi Systems, an artificial-intelligence research company based in Philadelphia that has done modeling work for the military, said, "Modeling emotional response to panic and fear is an area where everyone wants to go, but no one is really there yet."

The exercise simulating the Balkans mission is one step toward introducing emotional individuals into a situation so people can be trained for complex tasks like peacekeeping duties. "If the only interaction you have with an enemy is at the end of a gun," Dr. Rosenbloom said, "you don't need to know that much about them."

To enhance the realism, the Institute for Creative Technologies, with a $45 million grant from the Army, has built a theater here with a screen that wraps around roughly half the room. Three projectors and a sound system make the theater so realistic and directional that it can trick the listener into believing that a sound's source is coming from anywhere in the room.

After completing the exercise, a trainee receives an evaluation, said Dr. Bill Swartout, the institute's director of technology. "Depending on the path you took, a particular tape is played," he said. Because there are only a few possible paths in this version of the simulation, he said, it is possible to record the evaluations in advance.

As the simulation becomes more sophisticated, there will be more choices for the lieutenant, and software will put the story together on the fly.

Part of the problem with modeling emotions, Mr. Zachary said, is a dearth of experimental data in actual situations. "We just have anecdotes," he said, "because the ethics of research in modern American culture say you can't put people in experiments and scare them to death."

And few settings are more frightening than a battlefield. Even when there is information about actual behavior under dangerous conditions, it is linked so strongly to that particular time and place that few conclusions can be drawn. "You can have the best- laid plans, and once the flag goes up, those plans are gone," said Dr. David M. Nicol, a computer science professor at Dartmouth College. "Often it's the case that things hinge on the actions of one small group or one individual that shape the force or the outcome. So if you're dealing with statistical models of human behavior, you'd run into battle 10 times and get 10 different outcomes."

Modeling the behavior of one's own forces is hard, but building such models for adversaries or foreign civilians is much harder, even when their cultural and ideological foundations may be well documented.

One focus of Dr. Silverman's research, for instance, is terrorist behavior. Dr. Silverman evaluates his models by comparing them with the actual motives and behaviors observed for various groups.

In one project he is working on, the human player is the leader of a squad guarding a checkpoint at a bridge. All the other participants are simulations. In the exercise, a school bus approaches, filled with women and children. The bus also holds armed terrorists who are planning an attack.

"Throughout the ages," Dr. Silverman said, "we have been taught that emotions are the opposite of rationality and that cold logic is devoid of emotions." But new research shows that most decisions are guided by emotions, he said. "It's ironic, but to build realistic, clever software agents, we are giving them emotions and the capability of emotionally reacting to events and actions."

Dr. Silverman is optimistic about how quickly the new direction in research will prove effective. "It's definitely coming together," he said. "We're at the early stage, and there's a lot more theory than data. It's very easy to program a theory, but much harder to ground that in data and say this actually duplicates how people behave. But the field is moving rapidly forward."

But Dr. Pew of BBN Technologies is less optimistic: "I'm not sanguine that in the next five years we'll be there. People are complicated. You never do exactly the same thing twice. Every situation you face is always a little bit different, so trying to build a model that can reflect the importance of that context is where the challenges are."

To make further advances, Dr. Pew said, will require closer cooperation between psychologists and computer scientists. "If we want to be more successful with computer models, we need to go deeper into the psychology of how people perform," he said. "Because many of the models are ginned up by computer scientists who don't know anything about human behavior."