foundations of computational agents
Dimensions: flat, relational, infinite horizon, fully observable, deterministic, goal directed, non-learning, single agent, offline, perfect rationality
A robot that can deliver parcels to people needs a model of the world before it knows which parcels exist and which people may need deliveries. It might need to be programmed before it knows the environment which it will inhabit. A tutoring system needs to work for multiple students and multiple problems, and needs to be programmed before it knows about the students and all of the problems. A purchasing agent, when it is being designed and built, will not know about all of the hotels and rooms it can book, and will not know about the people and their goals or preferences. In all of these cases, the agent’s goals and its environment are described in terms of individuals and relations. When the agent’s knowledge base is built, and before the agent knows the objects it should reason about, it requires a representation that is independent of the individuals. Thus, it must go beyond feature-based representations. When the individuals become known, the agent may be able to ground the representations by substituting the known individuals for the logical variables, and just use features. Often, it is useful to reason in terms of the non-grounded representations.
With a relational representation, time can be reified, or made into an individual. Time can be represented in terms of individual points in time or temporal intervals. This section presents two relational representations that differ in how time is represented.