Knowledge Representation & Reasoning

Let us now turn to the issue of how to realize the agent function. In its simplest form, we can realize the agent function as a function that  maps a percept and directly map to the action to be executed next. This kind of realizing the agent function is called a reflex agent, which is often implemented in the form of simple if-then rules of the form "if you bump into an obstacle, turn around by 90 degrees.''

For robot agents that are to accomplish more complex tasks, a purely reactive action selection strategy is not sufficient. This holds for several reasons including that often environments are only partially observable, such that the best decision cannot be made on the current percept only. For example, consider the objects in a closed drawer, which the robot should remember after it has detected them. An effective approach to dealing with partial observability is to maintain an internal representation of the state of the environment, which is updated according to the percepts in every iteration of the perception action loop. For example, the state of the environment can include spatial relations of objects (e.g., locations and containers, etc.) as well as their internal state (e.g., switched on, fill level, etc.). We call this internal representation of the environment a model of the environment or a world model.

In many cases, the performance of the agent function can be further improved if the agent takes, in addition, its experiences and expectations into account. For complex tasks such as meal preparation, the action selection can be informed not only by the state of the environment but also based on the experience collected from former actions and the expectations about the effects of future actions. For example, in order to pick up an object from a drawer, the drawer must be opened first to make the objects inside accessible in the case that it is closed. In order to place an object at a specific location, it must be grasped with the agent's gripper first and in order to grasp an object, the gripper must be opened and must not already hold a different object. In order to take the right decisions, the agent thus must also have a model of its actions and their effects in the environment and which conditions the environment must be fulfilled for an action to be applicable. We refer to the specification of an action's effects and their applicability as an action's pre and post conditions. The preconditions represent conditions under which an action is executable and the postconditions specify the effects that the execution of an action has in the environment.

Fig. 1: Schema of a knowledge-enabled implementation of the rational agent function: The sequence of percepts is first interpreted by a perception component of the agent to get an updated formal representation of the world model. Based on the agent's actions and the world state, all possible action sequences are enumerated. By projecting forward the effects of each action in the sequence, a final state is obtained. If the final state satisfies the agent's goals, a plan for achieving a goal has been found, which is returned as a solution.

We consider a knowledge-enabled robot agent to be a robot agent that is equipped with symbolic representations of the rational robot agent model. An advantage of knowledge-enabled robot agents over reactive agents is that they can infer action sequences that are predicted to achieve a given goal. In this setting, the goal is being specified as the set of environment states that satisfy the given goal. For example, a possible goal specification of a popcorn preparation task is "the corn has popped and is located in a bowl. All devices are switched off.''

A possible realization of the process of inferring a goal-achieving action sequence a knowledge-enabled agent function is visualized in Figure 3. The action sequence generator enumerates all possible sequences of actions that the robot agent can perform, starting from the initial state of the environment model. By successively applying the pre- and postconditions of all actions in a sequence, the environment models' final state after execution of the whole sequence is obtained. If the resulting state satisfies a goal state, the agent function has found a solution and the respective action sequence is returned for execution. We call such an action sequence a robot action plan. If the goal state is not satisfied in the resulting environment state, the next action sequence from the action sequence generator is examined. The process of finding a sequence of actions that achieves goal states is called the planning problem.