Abstract
This chapter reports the last machine learning techniques for the assistance of automated planning. Recent discoveries in automated planning have opened the scope of planners, from toy problems to real-world applications, making new challenges come into focus. The planning community believes that machine learning can assist to address these new challenges. The chapter collects the last machine learning techniques for assisting automated planners classified in: techniques for the improvement of the planning search processes and techniques for the automatic definition of planning action models. For each technique, the chapter provides an in-depth analysis of their domain, advantages and disadvantages. Finally, the chapter draws the outline of the new promising avenues for research in learning for planning systems.
TopBackground
An AP task is defined by two elements: (1) a set of actions that represents the state-transition function of the world (the planning domain) and (2), a set of facts that represent the initial state together with the goals of the AP task (the planning problem). These two elements are typically represented in languages coming from the first-order logic. In the early days of AP, STRIPS was the most popular representation language. In 1998 the Planning Domain Definition Language (PDDL) was developed for the first International Planning Competition (IPC). Since that date, PDDL has become the standard representation language for the AP community. According to the PDDL specification (Fox & Long, 2003), an action in the planning domain is represented by: (1) the action preconditions, a list of predicates indicating the facts that must be true so the action becomes applicable and (2) the action effects, which is a list of predicates indicating the changes in the state after the action application. Like STRIPS, PDDL follows the closed world assumption to solve the frame problem. Regarding this assumption, what is not currently known to be true, is false.
Key Terms in this Chapter
Meta-Predicates: are extra predicates used to reason about the state of the search process in the planner, e.g., the goals that the planner is currently working on, the operators being considered, etc.
Problem Distribution: A set of problems belonging to a given domain generated with the same number of world objects and problem goals.
Bootstrap Effect: The fact of learning to solve a problem a teacher can not solve by observing how the teacher solves more simple problems.
Concept Language: also known as description logics, is a representation language with the expressive power of fragments of standard first-order logic but with a syntax that is suited for representing and reasoning with classes of objects.
Utility Problem: It is a drawback that arise when using learned knowledge cost of using overwhelms its benefit because the difficulty of storage and management the learned information and because of determining which information to use to solve a particular problem.
Closed World Assumption: Any formula non explicitly asserted in a state is taken to be false, which allows one to avoid the explicit specification of negated literals. This assumptions presupposes that actions only change a small part of the world.
The Frame Problem: The problem of expressing a dynamic system using logic without explicitly specifying which conditions are not affected by an action.
Model Checking: consist of determining whether a given property, usually described as a temporal logic formula, holds in a given model of a system. Traditionally, this problem has been studied for the automatic verification of hardware circuits and network protocols.