Planning (artificial intelligence)

The planning ( Automated planning ) is a discipline of the Artificial intelligence which aims at the development of algorithms to produce plans (in other terms, a Planification), typically for the execution by a robot or any other agent. The software of planning which incorporates these algorithms names planning .

A typical planner handles takes three entries (all coded in a formal language such as STRIPS which uses Prédicat S logical):

a description of the initial state of a world,
a description of a goal to reach and
a whole of possible actions (sometimes called operators)

Each action generally specifies preconditions which must be present in the actual position so that it can be applied, and of the postconditions (effects on the actual position).

The difficulty of the problem of planning depends on the assumptions of simplification which one takes for asset, for example an atomic time, a deterministic time, a complete observability, etc

The traditional problem of planning

The traditional planning take for asset that all these assumptions hold. They were studied in-depth. Some popular techniques are

front research in a space of states,
back research in a space of states,
front research in a space of plans, Graphplan, and
the transformation towards a problem of satisfiability of proposals.

The Algorithme A* is a typical example of traditional algorithm of planning, often employed in the courses of introduction for its simplicity.

Practical planning

In practice, one vérife only very seldom assumptions of traditional planning. This is why good number of extensions were born.

The nondeterministic problem of planning

If the assumption of the determinism is given up and a probabilistic model of uncertainty is adopted, then this leads to the problem of the generation of policy (or strategy) for a Markovian Decision-making process (MDP) or (in the general case) a Decision-making process Markovian partially observable (POMDP).

Nonlinear planning

Traditional planning résoud under-goals in a given order. The problem is sometimes that brings to destroy what was already built. This phenomenon is known under the name of anomaly of Sussman.

Let us suppose that an individual barefeet must find himself in the state or it carries his right shoe, its left shoe, its right sock and its left sock. If he seeks to achieve the goals in the order of the statement, he will fail.

To solve this type of problem, one can pass in partially ordered plans in which the order between the actions is fixed only when it is necessary ( engagement at the latest or least commitment planning ).

In the preceding example, to put the left shoe must be made after having put the left sock. Idem for the line. On the other hand the execution of the plan for the left is independent of the execution for the line. The global level is thus partially ordered.

The planners able to manage this category of problem are said to partial order (POP, NOAH, etc).

Hierarchical planning

Certain problems are not easily soluble in the state because of their complexity. One can gain in effectiveness by carrying out an abstraction of the details and by changing the granularity of the operators of planning. One can for example start to plan with high level then to go down in detail to the need (like ABSTRIPS does it for example).

See too

Random links: