Environment definition

Now, after developing the definition of rationality, we are almost ready to start creating rational agents. But first, it is necessary to determine what the problem environment is, which is essentially a “problem” for which a rational agent serves as a “solution”.

We begin with a demonstration of how to identify a problem environment, and illustrate this process with a number of examples. Then in this section it will be shown that the problem environment can have a variety of varieties. The choice of the project that is most appropriate for a particular agent program directly depends on the type of problem environment under consideration.

Problem Environment Definition

In the above study of the rationality of a simple agent vacuum cleaner, we had to determine the performance indicators, the environment, as well as the actuators and sensors of the agent. Group the description of all these factors under the heading Problem Environment. For those who love abbreviations, the authors abbreviated the corresponding description as PEAS (Performance, Environment, Actuators, Sensors — performance, environment, actuators, sensors).

The first stage of the design of any agent should always be to identify the problem environment with the greatest possible completeness. The example in which the world of the vacuum cleaner was considered was simple; Now consider a more complex problem - the creation of an automated taxi driver.

This example will be used throughout the remainder of this chapter. Before the reader feels anxiety about the safety of future passengers, we want to immediately note that the task of creating a fully automated taxi driver is currently still beyond the capabilities of the existing technology.

A complete solution to the problem of driving a car is extremely time consuming, since there is no limit to the appearance of more and more new combinations of circumstances that may arise in the process of driving; This is another reason why we chose this issue for discussion. The table provides a summary description of PEAS for a problem taxi driving environment.

Agent type	Performance indicators	Wednesday	Actuators	Sensors
Taxi driver	Safe, fast, comfortable ride within the rules of the road, profit maximization	Roads, other vehicles, pedestrians, customers	Steering, accelerator, brake, lights, horn, display	Video cameras, ultrasonic range finder, speedometer, global navigation and positioning system, odometer, accelerometer, engine sensors, keyboard

First of all, it is necessary to determine the performance indicators with which we could stimulate the activity of our automated driver. Desired qualities include the successful achievement of the desired destination; minimization of fuel consumption, wear and aging; minimizing the duration and / or cost of the trip; minimizing the number of traffic violations and interference with other drivers; maximizing the safety and comfort of passengers; profit maximization. Of course, some of these goals are in conflict, so possible compromises should be considered.

Then consider what is the driving environment in which the taxi operates. Any taxi driver has to deal with a wide variety of roads, starting from country lanes and narrow city lanes and ending with freeways with twelve lanes. On the road, there are other vehicles, street animals, pedestrians, road workers, police cars, puddles and potholes. The taxi driver also has to deal with potential and actual passengers. In addition, there are several other important additional factors.

A taxi driver may have a chance to work in Southern California, where a problem such as snow rarely arises, or in Alaska, where there is no rare snow on the roads. It may turn out that the driver will have to drive on the right side all his life, or it may be necessary for him to be able to adapt quite successfully to the driving on the left side during his stay in Britain or Japan. Of course, the more limited the environment, the simpler the design task.

The actuators available in an automated taxi should be more or less the same as those at the disposal of a human driver: engine controls using an accelerator and driving controls using a steering wheel and brakes. In addition, it may require means of displaying or synthesizing speech on the screen for transmitting response messages to passengers and, possibly, certain ways of communicating with drivers of other vehicles, sometimes polite and sometimes not quite.

To achieve their goals in this driving environment, the taxi driver will need to know where he is, who else is driving along this road and with what speed he is moving. Therefore, its main sensors should include one or more controlled television cameras, a speedometer and an odometer. For proper driving, especially when cornering, an accelerometer must be provided; the driver will also need to know the mechanical condition of the car, so it will need a conventional set of sensors for the engine and electrical system.

An automated driver may also have instruments that are unavailable to the average human driver: satellite global positioning and navigation system (Global Positioning System — GPS) for obtaining accurate position information with respect to the electronic map, as well as infrared or ultrasonic sensors for measuring distances to other cars and obstacles. Finally, he will need a keyboard or microphone for passengers so that they can indicate their destination.

The table briefly lists the main elements of PEAS for a number of other types of agents. It may seem surprising to some readers that the authors included in this list of agent types some programs that function in a completely artificial environment, limited by keyboard input and character output to the screen. Some might say: “Of course, this is not a real environment, is it?” In fact, the point is not in the differences between “real” and “artificial” environmental variants, but in the complexity of the links between the behavior of the agent, sequence of acts of perception, produced by this environment, and performance indicators.

Some “real” media options are actually extremely simple. For example, for a robot designed to control parts passing past it on a conveyor belt, a number of simplifying assumptions can be used, for example, that the lighting is always on, that the only items on the conveyor belt are parts of the type that he knows, and that only two actions (to accept the product or reject it).

Agent type	Performance indicators	Wednesday	Actuators	Sensors
Medical Diagnostic System	Successful recovery of the patient, minimization of costs, lack of reasons for litigation	Patient, hospital, staff	Conclusion of questions, tests, diagnoses, recommendations, directions	Keyboard typing of symptoms, lab results, patient responses
Satellite Image Analysis System	Proper image classification	Data channel from the orbital satellite	Displaying the classification results of a certain fragment of the image	Pixel arrays with color data
Robot sorter details	The percentage of error-free sorting by tray	Belt conveyor with moving parts; trays	Articulated arm and grip	Video camera, angle sensors of hinges
Cleaning plant controller	Maximizing the degree of purity, productivity, safety	Cleaning plant operators	Valves, pumps, heaters, displays	Temperature, pressure, chemical composition sensors
Interactive English Language Training	Maximize student grades on exams	Many students, an examination agency	Displaying exercises, recommendations, corrections	Keyboard input

In contrast, some software agents (also called software robots or softbots) exist in complex, unbounded problem areas. Imagine a software robot designed to control a simulator simulating a large passenger plane. This simulator is a very detailed modeled, complex environment in which the movements of other aircraft and the work of ground services are simulated, and the software agent must choose the most appropriate actions in real time from a wide range of actions.

Another example is a software robot designed to view news sources on the Internet and show customers messages that interest them. To succeed, he needs certain abilities to process text in natural language, he must determine what interests each customer in the learning process, and also be able to change his plans dynamically, for example, when a connection to any news source closes or operatively mode goes new news source. The Internet is an environment that, by its complexity, competes with the physical world, and many artificial agents are among the inhabitants of this network.