Multi-agent systems consist of agents and the environments where they operate. Agent environments can be categorized along various traits, but the most cited is probably the classification presented by Russell and Norvig.
They organize the environments according to the following properties:
Accessible vs. inaccessible – if it is possible to gather full and complete information about the environment in the moment, then the environment is accessible. Typically, only virtual environments can be accessible, because in reality, all sensors provide an input that is biased and incomplete up to some extent. There are so many potential percepts in the real world that it would be impossible to record and process them in real time (even if agent’s sensors were infinitely sensitive). In fully accessible environments, the agents do not need to create models of the world in their memories, because it can get any needed information from the environment at any time.
Deterministic vs. non-deterministic – if an action performed in the environment causes a definite effect, the environment is deterministic. Definite effect means that any action of the agent leads to the intended and expected results and there is no room for uncertainty. Of course, if the environment is inaccessible for the agent, it will be probably non-deterministic, at least from its point of view. Turn-based games are an example of a typical deterministic environment, whereas a room with a thermostat (where the thermostat is the agent) is an example of nondeterministic environment, because the action of the thermostat does not necessarily lead to the change of temperature (if, for instance, a window is open).
Static vs. dynamic – the environment is static when the agent is the only entity that changes the environment in the moment. If it changes during the agent’s action (i.e., the state of the environment is contingent on time), it is dynamic. Again, often real environments are dynamic (e.g., traffic in a city) and just some artificial environments are static (consider turn-based games like chess again).
Discrete vs. continuous – this depends on whether a number of possible actions in the environment are finite or infinite. If the agent just has a certain set of possible actions that it can do in the moment, then the environment is discrete. Otherwise, when the agent has theoretically an infinite number of options, the environment is continuous. Suppose that roulette is a discrete environment. The agent can place a wager on a certain, limited number of betting areas. On the other hand, the legal system is a continuous environment. People have theoretically an unlimited number of options how to, for example, close deals or defend themselves before a court.
Episodic vs. non-episodic – episodic environment is the environment where the agent operates in certain segments (episodes) that are independent of each other. The agent’s state in one episode has no impact on its state in another one. Human life exists in a non-episodic environment, because all of our past experiences influence our conduct in the future. An operating system, on the other hand, is an episodic environment, as we can reinstall it. Then programs-agents can be installed on a “clean system” with no conjunction with the same programs installed on the old system.
We can distinguish the environments also according to their spatial characteristics. It can be particularly useful in the case of agent-based models:
Dimensional vs. dimensionless – if spatial characteristics are important factors of the environment and the agent considers space in its decision making, then the environment is dimensional. If the agents do not take space into account, then the environment is dimensionless. Real environments are typically dimensional, as we naturally feel and count with spatial characteristics of our surroundings. In the virtual environments, such characteristics are not always important. For example, as stock markets are almost fully electronic today, it does not matter where somebody is present physically, since he or she can buy or sell shares on theoretically any market in the world. In such an environment, spatial characteristics have no influence on the agents’ decision making, and therefore it is dimensionless.