| { | |
| "paper_id": "T75-1003", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T07:43:20.352870Z" | |
| }, | |
| "title": "AUTOMATIC PLANNING FROM A FRAMES POINT OF VIEW", | |
| "authors": [], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "", | |
| "pdf_parse": { | |
| "paper_id": "T75-1003", | |
| "_pdf_hash": "", | |
| "abstract": [], | |
| "body_text": [ | |
| { | |
| "text": "In this note I would like to consider some of the ways in which Minsky's ideas concerning frames relate to automatic planning systems. I will proceed by making some observations about how knowledge is represented in planners and will relate some of the frame ideas to actual running systems.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Hopefully, the presentation will increase the reader's understanding and intuition of both the systems and the ideas.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Typically, a planning system is one that assists in the accomplishment of tasks by generating plans and then monitoring their execution. The active agent that actually carries out the plans might be another part of the system (e.g. a mechanical manipulator), or a human that the system is providing instructions to.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "A task is usually specified to such a system by describing an initial situation and a desired (goal) situation.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The system is aware of a collection of actions that the active agent can carry out, and the plan it produces is a sequence of these actions that are expected to transform the initial situation into the desired situation.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Taking a broader view, a planner can be thought of as being a constructor of scenarios or thematic frames for whatever reasons they are needed. For example, the system may be involved in understanding a natural language narrative that describes a sequence of actions, or it may be participating in an instructional dialogue with a human about some procedure, or it may be answering questions about how some task would be accomplished.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The building blocks that a planner uses to construct a new plan can be usefully thought of as frames that describe some action.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "II. Action Frames as Plan Steps", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "In order for a planner to make use of such existing \"action\" frames, they must include certain descriptive information about themselves. In addition, for the new frame to be useful in situations other than just reoccurrences of the one for which it was originally constructed, it must somehow be generalized so that, like the other action frames, there are slots at the frame's \"terminals\" that can be filled in \u2022 with values suited to any of a class of situations and goals.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "II. Action Frames as Plan Steps", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The basic goal of the STRIPS learning work was to generalize and save plans so that they could be used as single steps in fUture plans.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "II. Action Frames as Plan Steps", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Two aspects of this work seem worth mentioning here.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "II. Action Frames as Plan Steps", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "First, the preconditions for the plan were directly available as a side effect of the determination of \"kernels\" preceding each step of the plan.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "A. Kernels", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The kernel preceding step i of a plan is a partial description of the situation that the planner expects to exist after execution of the first i-I steps of the plan. Hence, the planner can add any number of steps between step i-I and step i of an existing plan as long as the kernel between those two steps is not violated.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "A. Kernels", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The second comment to be made about the STRIPS learning work relates to the generalization that was done on the plans before they were stored. The goal was to \"unbind\" the slots in each of the plan's steps so that, whenever possible, they became unvalued slots in the frame for the entire plan.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "H. Plan Generalization", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The logical structure of the plan imposes restrictions on this unbinding process so that some pairs of slots are required to take on the same value (i.e., they become a single slot in the new frame), and others must retain their binding (i.e., they lose their status as slots in the new frame).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "H. Plan Generalization", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "For example, if one step of the plan causes the robot to go to a door, and the next step causes the robot to go through a door, then the door in those two steps must be the same door.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "H. Plan Generalization", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "This generalization process essentially parameterizes the plan and thereby provides it with enough generality to make it worthwhile to save.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "H. Plan Generalization", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "For example, consider a plan that takes a robot from room RI through aoor DI into room R2 and then has the robot bring box BI from room R2 back into room ~I through door DI. That plan would be generalized so that it would take the robot z rom any room into any adjacent room through any connecting door and then take any box in the room into any adjacent room through any connecting door.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "H. Plan Generalization", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The generalization removes all bindings to the particular rooms, doors, and box, and allows the room into which the box is taken to be different from the one in which the robot was initially located. [Fahlman, 1974] In that work, we wish to allow a manager to specify to the system various operational procedures that he uses in his organization.", | |
| "cite_spans": [ | |
| { | |
| "start": 200, | |
| "end": 215, | |
| "text": "[Fahlman, 1974]", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "H. Plan Generalization", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The system can then act as an administrative assistant by planning and monitoring the execution of these procedures at the manager's request. The planning activity in this system is primarily one of scheduling the individual steps of an operation, and this scheduling is concerned with assuring the availability of resourses and personnel at specific times.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "H. Plan Generalization", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "In particular, we have been considering management problems on board a Navy aircraft carrier.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "H. Plan Generalization", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "One operation of interest in that domain is the flying of training missions. Such a mission involves steps such as preflight and postflight maintenance of the aircraft, fueling the aircraft, briefing and ~ebriefing the pilot, launching and recovering the aircraft, etc. These steps and the order in which they must occur do not vary from mission to mission, hence they can be included in a thematic frame that is part of the definition of a mission.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "H. Plan Generalization", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The mission scenario or action frame has many terminals with unassigned values, and the planner's basic task is to find an acceptable set of values for them. Most of these slots specify the start or the end time of some step in the plan; others specify the identity of the pilot and aircraft, the amount of fuel to be carried, bearings and locations to indicate the flight route, ere.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "H. Plan Generalization", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "We include at each terminal of the mission's action frame a set of constraints on the value to be assigned there.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "H. Plan Generalization", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The most common constraints are those that are derived from the temporal partial ordering of the steps (e.g.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "H. Plan Generalization", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "fueling of the aircraft must occur after the preflight maintenance and before the launch). Default values are included at the terminals, usually expressed as functions of other slot values in the frame.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "H. Plan Generalization", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "For example, the default value for the start of preflight mainten@nce is expressed as the time of launch minus constant.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "H. Plan Generalization", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "These default values allow the planner to make feasibility estimates and to make scheduling decisions before all of the constraints from the other schedulers have oeen determined.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "H. Plan Generalization", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Also included at each terminal is a specification of what part of the system can be called upon to determine a value for the terminal.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "H. Plan Generalization", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "These ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "H. Plan Generalization", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Even though subplanners are procedural in nature, they can usefully be embedded in frame-like structures as if they were actions ~o be included as single steps in a plan.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "V. Subplanners as Action Frames", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "That is, we can talk about the preconditions, effects, task relevance, etc.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "V. Subplanners as Action Frames", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "of subplanners such as the tower builder, the route tinder, and the device assembler as if they were single actions. These The availability of such meta-actlon frames allows the system to do \"hierarchical planning,, in the sense that a meta-actlon can be included in a plan without calling its subplanner to determine the plan steps that it represents. ~974,1975] , and found to PrOVide significant advantages both during the generation and the execution of Plans. ~n these systems, plans tend to grow in a", | |
| "cite_spans": [ | |
| { | |
| "start": 353, | |
| "end": 363, | |
| "text": "~974,1975]", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "V. Subplanners as Action Frames", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\"breadth-first, manner in that typically a complete plan will be constructed using nigh level meta-actions %efore the detailed steps of any of the meta-actions are determined. Since a meta-action's subplanner may itself Produce a plan containing meta-actions, a multiple level plan hierarchy can be formed, and the planner can make independent decisions for each meta-action as to whether its subplanner should be called.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "top-down", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The advantages of hierarchical planning derive basically from the fact that by using meta-aetions the system can ignore detailed actions until it is confident that they will become part of the final plan.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "top-down", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Indeed, in many situations the detailed steps may not be needed at all. For example, the system may be giving instructions to a human who is skilled in the task domain and therefore does not need detailed instructions; or the system may be only answering a specific question about a task such as \"Can the robot fetch the box?\" or \"How long will it take to replace the pump?\".", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "top-down", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [ | |
| { | |
| "text": "Ii.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "annex", | |
| "sec_num": null | |
| } | |
| ], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "A Planning System for Hobot Construction Tasks", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [ | |
| "E" | |
| ], | |
| "last": "Fahlman", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": null, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Fahlman, S. E., \"A Planning System for Hobot Construction Tasks\",", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "num": null, | |
| "uris": null, | |
| "type_str": "figure", | |
| "text": "whenever the scenario needs to be made explicit." | |
| } | |
| } | |
| } | |
| } |