PRACTICAL PLANNING EXTENDING THE CLASSICAL AI PLANNING PARADIGM

PRACTICAL PLANNING: EXTENDING THE CLASSICAL AI PLANNING PARADIGM DAVID E. WILKINS ARTIFICIAL INTELLIGENCE CENTER SRI INTERNATIONAL Tarkan EYERCİ Sedat AKEL

About the Book ◦ Planning, or reasoning about actions, is a fundamental element of intelligent behavior--and one that artificial intelligence has found very difficult to implement. The most well-understood approach to building planning systems has been under refinement since the late 1960 s. ◦ Practical Planning is an in-depth examination of this classical planning paradigm through an intensive case study of SIPE, a significantly implemented planning system. The author, the developer of SIPE, defines the planning problem in general, explains why reasoning about actions is so complex, and describes all parts of the SIPE system and the algorithms needed to achieve efficiency. Details are discussed in the context of problems and important issues in building a practical planner; discussions of how other systems address these issues are also included.

Reasoning About Actions ◦ Reasoning about actions is an essential component of intelligent behavior, and an essential part of future intelligent computer systems. ◦ But this is very difficult to formalize in a computerized environment. ◦ Reasoning about the effects of actions can solve practical problems in business as well as the military ◦ Accurate planning and forecasting can keep projects on schedule and prevent cost overruns. ◦ Forseeing the future implications of business decisions can help companies to avoid mistakes. ◦ Planning faster and more completely than an opponent can favorably decide the outcome of a battle or a war

Reasoning About Actions in AI ◦ Generating a plan of action for a robot, ◦ Interpreting an utterance by reasoning about the goals a speaker is trying to achieve, ◦ Automatically writing a program to solve a problem, ◦ Producing utterances in a conversation with another agent, ◦ Allocating the use of resources, computational or otherwise.

Reasoning Robot ◦ A robot capable of intelligently planning actions could potentially have a large economic impact. ◦ Reasoning robots could function in conditions that are dangerous to humans ◦ ◦ Building a space station in a vacuum, Cleaning up toxic or radioactive waste, Constructing or retrieving underwater objects, Accomplishing missions on a battlefield. ◦ Robots would also be valuable in safer environments. ◦ Running household errands, ◦ Accomplishing storage and retrieval tasks in a warehouse

The Classical AI Planning Problem ◦ Research on reasoning about actions is a subfield of AI labeled "planning. " ◦ The first attempts at AI planning, such as the STRIPS and NOAH systems developed at SRI International during the late sixties and early seventies ◦ The classical definition of the planning problem assumes a state-based representation of the world. ◦ Alternatives to state-based approach are procedural representations, and event-based representations.

The Classical Planning Problem

Reactive Planning ◦ None of the classical AI planners are reactive ◦ Reactive autonomous agents must be able to plan their actions in advance ◦ Some argue that a distinction between planning and execution is wrong, and that we do not need planning, only reactive controllers ◦ Author stated that «a robot should have both a planning component and a reactive control system» ◦ Reasoning about actions is necessary for the robot to act in accordance with a long-term strategy for achieving its overall goals

The Essence of Planning ◦ A system can reason about the world both as it was before the event and as it will be after the event. This capability is necessary if a computer system is to plan a course of action. ◦ This problem is known as the frame problem. ◦ State-based representations, involves the enormous number of facts about the world that do not change when an event occurs (e. g. , driving to work does not change the color of the sky or the net worth of IBM) ◦ Another difficult aspect of the frame problem is that the facts that do change when an action is taken depend on the exact situation in which the action is performed. ◦ For example, driving your car to work at 8 a. m. will affect the world differently than driving the car to work at 10 a. m. , assuming that the trip will take longer during the 8 o'clock rush hour. ◦ Thus the effects of an action depend on the particular situation (context) in which an action is executed. We refer to such effects as context-dependent effects. ◦ Reasoning about how actions affect the world is the heart of the planning problem.

Capabilities of a Planning System

Capabilities of a Planning System ◦ Causal theory ◦ Deducing context-dependent effects to solve the frame problem ◦ Necessary for sufficiently expressive operators ◦ Nonlinear ◦ if it contains actions that are unordered with respect to each other, i. e. , actions for which the planner has not yet determined an order and which may possibly be in parallel. If a planner has the ability to represent and reason about nonlinear plans, it can avoid committing to a particular order of action until information for ordering the actions has been accumulated. This can avoid an exponential search of all possible plan orderings. ◦ Allowing nonlinear plans can pose problems for basic planning algorithms. ◦ Two actions interfere with each other if they are executed at the same time, and a nonlinear planner must recognize and correct these situations in order to generate correct plans. ◦ Avoids search of all possible plan orderings

Capabilities of a Planning System ◦ Hierarchical planning ◦ ◦ In complex domains it is crucial to plan at different levels of abstraction. This is an abstract level of reasoning If a planner began planning at too detailed level, it would be overwhelmed Avoids search of all detailed plans ◦ Planning variables ◦ Allow certain objects or entities within a plan to be left unspecified during the planning process ◦ For example, the planner can plan to go somewhere without initially specifying exactly where by representing the intended location as a planning variable ◦ The use of planning variables eliminates a huge search space since later information may determine the correct instantiation for a variable without the system having to produce and check a plan for every possibility. ◦ Avoids search of every plan instantiation

Capabilities of a Planning System ◦ Constraints ◦ The search through possible instantiations for variables can be further reduced if the planner can flexibly narrow the set of possible instantiations for a variable ◦ if the user can specify that the location must be within 100 miles of San Francisco and yet must not be San Jose, then the set of possible locations has been narrowed substantially. ◦ Such limitations on the instantiations of a variable are called constraints. ◦ Further avoids search of plan instantiations ◦ Replanning ◦ When plans are executed in the real world, events rarely proceed exactly as planned. If a planner can patch its original plan after an unexpected event and continue using the modified plan, it can potentially save considerable effort over planning again from scratch. ◦ Avoids restarting after an unexpected occurrence

Capabilities of a Planning System ◦ Domain-independent planning system provides is a representation that is particularly geared towards representing actions and reasoning about how their effects change the state of the world. ◦ By forcing the domain knowledge to be encoded in its formalism, the planner is able to use its solutions to the frame problem on whatever domain the user chooses to encode. ◦ Many domain-specific "planning" systems avoid the need to reimplement the types of solutions in this code by formulating their problem in such a way that reasoning about actions in a general way is avoided ◦ Allows solution of broad range of problems

How Hard is Planning? ◦ The heart of a planning system, from the standpoint of efficiency, is its truth criterion ◦ Nonlinear plans contain unordered actions. ◦ A sufficiently expressive planning formalism requires the deduction of context-dependent effects ◦ So many combinatorial problems to be addressed, the progress would be slow ◦ One of the primary aims of the classical AI planning paradigm is to address this complexity

Classical AI Planning Systems ◦ The classical AI planning paradigm provides four of the features in the table. ◦ Aclassical AI planning system ◦ ◦ domain-independent, permits planning variables, supports hierarchical planning, permits nonlinear plans.

Features of Existing Systems