V. Richard Benjamins1 - Leliane Nunes de Barros2 - André Valente3
1University of Amsterdam, Social Science Informatics (SWI), Roetersstraat 15, NL-1018 WB Amsterdam, The Netherlands, firstname.lastname@example.org
2University of Maryland, Department of Computer Science, College Park, MD 20742, USA, email@example.com
3University of Southern California, Information Sciences Institute, 4676 Admiralty Way, Marina del Rey, CA 90292, USA, firstname.lastname@example.org
When engineering a planning application, a serious knowledge engineer would believe it is wise to search the literature on planning looking for a method that is adequate for the problem at hand. He/she would, however, probably be overwhelmed by the enormous amount of literature and the almost total absence of a clear synthesis of the available methods and their applicability. Most of the methods are described in terms of search, which offers little help in understanding the applicability of the method to a certain domain. The few analyses of the existing methods have concentrated on characteristics such as efficiency, soundness and completeness, and expressiveness --- which, again, do not help to construct an application. Clearly, something is necessary to ``make sense'' of the methods available and help in the knowledge engineering and acquisition process. A first and essential step towards developing tools to automate knowledge acquisition for planners is to have a model of what existing planning methods do and how they work [Kingston et al., 1996].
In this paper, we present a knowledge-level framework for representing and analyzing planning methods that can help knowledge engineers in constructing planning applications. The result of the analysis can be viewed as a library of problem-solving methods for planning, and as such we will use our analysis for knowledge acquisition and engineering. The library specifies reusable components to configure problem solvers rather than to build them from scratch. These components are represented, based on the knowledge-level framework of the CommonKADS methodology [Schreiber et al., 1994]. The library is based on three interconnected groups of building blocks. The first group of components is used to specify how domain knowledge is structured by the methods, the second is used to specify how the methods actually solve a planning problem, and the third is used for connecting these internally and to an application domain. The components are:
Earlier versions of (parts of) the library presented in this paper were discussed in [de Barros et al., 1996,Valente, 1995]. In [Valente, 1995], we discussed in detail the knowledge roles and domain models. In [de Barros et al., 1996] we concentrated on the basic methods and the task-method decomposition structure. In this paper, we focus on the knowledge acquisition benefits of the library, by concentrating on the third group of components, namely the assumptions. We also extend the library in order to cover a broader class of planning problems.
The aim of this paper is to show how the proposed library can help to construct planners for particular application domains by providing a high-level, abstract synthesis of the available planning methods. Specifically, we will concentrate on an analysis of the so-called assumptions of problem-solving methods as the basic connection between reasoning and domain knowledge. We will show how these features can be used both to (i) characterize a method and find out to which domains it is applicable, and (ii) characterize a domain and find out which methods are adequate to be used. In order to partially automate this process, we stored the problem-solving components in a tool called TINA ( Tool in Acquisition). Currently, the library concentrates on so-called classical planners, which implies that it inherits their limitations and behavioral assumptions. For example, it assumes there is a single agent; changes in the world can only be caused by the agent; actions do not take time to be performed; etc. The library can, however, naturally be extended to comprise other types of planners (e.g., case-based, planners that use uncertainty) as well as methods for other types of planning problem (e.g., plan execution, plan recognition).
This paper is structured as follows. In Section 2, we briefly describe the CommonKADS framework. Section 3 summarizes the domain part of the library, namely the knowledge roles and domain models. Section 4 describes the task part of the library, namely the task-method decomposition tree and the basic methods used to compose it. In Section 5, we describe how our library can support the knowledge acquisition process. Two kinds of support are demonstrated: (1) given a description of domain characteristics, to find what possible planners are adequate for the domain, and (2) given a planner, define its assumptions (domain requirements). At the end of the section, we discuss the merits and limitations of the approach followed, and, in Section 6, we present conclusions.
CommonKADS is a methodology for developing knowledge-based systems. The components used to represent problem-solving knowledge include the following:
Tasks are represented using a functional perspective, i.e., specified by a set of input/output relations which is called a function. A function includes a goal, i.e., a specification of what needs to be achieved. For example, the goal of a planning task is to come up with an ordered sequence of operations that transforms a given initial state into a goal state. We represent the input and output by roles that knowledge plays in the problem-solving process. For instance, in diagnosis, an input or output role like hypothesis means that some domain structure or element plays the role of hypothesis. Static roles are roles whose contents do not change during problem solving (they correspond roughly to a knowledge base in a knowledge-based system); dynamic roles are roles whose contents change.
Problem-solving methods (PSMs) describe how the goal of a task can be achieved. A method has input and output roles, and either decomposes a task into subtasks (composite method) or is primitive. Tasks and PSMs can be organized in a task-method decomposition structure, where a task may be realized by alternative methods, each of which consists of subtasks or is primitive. Figure 2 presents such a diagram for planning. A primitive method can usually be implemented by a general or weak method (e.g., constraint propagation). A composite method specifies the allowed data flow (in terms of their roles) between the subtasks, and the control knowledge over them. Control knowledge is represented by a control structure, which determines the execution order and iteration of the subtasks. This paper does not elaborate on control knowledge.
Assumptions are important constructs to make explicit the interaction between a PSM and the domain knowledge, as demonstrated by recent work [Benjamins & Pierret-Golbreich, 1996,Benjamins et al., 1996]. In the context of this paper, we will use assumptions to specify the requirements of a PSM on domain knowledge, which corresponds to the ontological assumptions of [Benjamins et al., 1996] and the epistemological assumptions of [Benjamins & Pierret-Golbreich, 1996]. Epistemological assumptions are further divided into availability assumptions, which refer to the availability of domain knowledge, such as the availability of a causal model in diagnosis; and property assumptions, which refer to properties of the domain knowledge, for instance, that a causal model is non-cyclic.
Domain ontologies provide a vocabulary that can be used to describe a certain domain. An ontology specifies the concepts, relations, etc. independently from the specific system that is reasoned about. Examples of domain ontologies in planning domains are the formalisms used to represent the world, time and actions (e.g., the situation calculus or the event calculus). We will not deal further with domain ontologies in this paper. Domain models are structures to be used in problem-solving methods, built using elements of the domain ontology. Typical examples are models that specify causal dependencies between symptoms and diseases in medical domains (causal models), or models that specify structural hierarchies for physical devices (structural models).
One of the critical elements in the analysis of a planning method is to specify which types of knowledge it employs --- in other words, which roles knowledge can play. In this section, we will present (i) a set of four general dynamic roles in planning tasks, and (ii) a part-of organization of static roles, originally defined in [Valente, 1995]. These two sets give a high-level view of the types of knowledge used in planning, both in the sense of the knowledge that is input or output to the task, and in the sense of the knowledge that used as part of the knowledge base.
The basic problem solved in planning tasks is to generate a sequence of actions (a plan) whose execution accomplishes a given goal. A planning problem is characterized by two inputs: an initial state and a goal description of the world. We distinguish the following dynamic roles in planning PSMs:
Figure 1 shows a part-of organization of static roles [Valente, 1995]. The fact that this is a part-of tree means that every planning method must use as a static role either the root or one of the logical ANDs of its constituting decompositions.
Figure 1: Part-of tree of static roles in planning.
The plan model role defines what a plan is and what it is made of. It consists of two parts: the world description and the plan description. Below is a brief description of these roles and their sub-roles.
The world description role describes the world about which we are planning and comprises two sub-roles: state description and state changes. The state description contains the knowledge necessary to represent or describe the state of the world. Examples of domain models to play this role in planning systems are a set of first-order predicates (STRIPS-like) or a set of fluents from the Situation Calculus [McCarthy & Hayes, 1969]. The state changes role comprehends all the information connected to the specification of changes in the state of the world. This is also the specification of the elements a plan is composed of (but not how they are composed, see plan composition below). Examples of domain models to play this role in planning systems are: the add-delete lists (ADL) used in STRIPS [Fikes & Nilsson, 1971]; or the hierarchical task networks (HTNs) [Tate, 1977,Erol et al., 1994] which are, in fact, descriptions of sub-plans; or skeletal plans [Friedland & Iwasaki, 1985] which are abstract descriptions of previously generated plans. In case of HTNs, the sub-plans may come fully or partially specified. Fully specified means that the sub-plans contain all the auxiliary constraints necessary for continuing the planning process (e.g., ordering constraints or truth constraints such as causal-links). Partially specified means that the sub-plans come with a correct and useful task decomposition but not with all the needed causal-links (for a real world planning application).
The plan description role describes the structure and features of the plan being generated and comprises two sub-roles: plan structure and the (optional) plan assessment knowledge:
(1) Plan structure This role specifies how the parts of a plan (actions, sub-plans) are assembled together. It also specifies (indirectly) how the plan is to be executed. There are several varieties in the structure of plans that can be identified in the literature. They can be described by two main knowledge roles: the plan composition role contains the description of the plan with respect to how the state changes are arranged in order to make up a plan. This includes, for instance, whether the plan will be a partial or a total ordering of a set of state changes, or whether it includes iteration or conditional operators. The composition may also be hierarchical: plans are composed of sub-plans, and so on up to atomic plans, which are normally state changes. The state change data role contains the plan information besides the structure of state changes. For example, important state change data are interval constraints for binding the variables involved in the state changes. It is also possible to assign different resources to each state change or sub-plan. Two particularly important resources are agents and time.
(2) Plan Assessment Knowledge determines whether a certain plan (or sub-plan) is valid (hard assessment knowledge), or whether a plan is better than another (soft). Based on this knowledge, a plan can be modified or criticized. An example of hard plan assessment knowledge is the truth-criterion, which is used to find out if a condition is true at some point in the plan. TWEAK [Chapman, 1987] uses a modal truth criterion (MTC) which defines when a particular condition in a plan is necessarily or possibly true by formally stating all possible interaction problems. Another example of a domain model for hard plan assessment knowledge is causal-link knowledge used in SNLP [McAllester & Rosenblitt, 1991]. Knowledge for testing the consistency of the auxiliary constraints is also an example of hard plan assessment knowledge. An example of soft knowledge would be that plans with smaller number of plan-steps are preferred. This paper only deals with hard assessment knowledge. Assessment knowledge can also be domain-specific, in which case it can be used to criticize the plan more efficiently.
Based on an analysis of many classical planning systems, we have identified relevant tasks and problem-solving methods. They can be organized into a task-method decomposition structure [Orsvärn, 1996] (see Figure 2), where a method consists of (solid lines) subtasks and a (sub)task can be realized by alternative (dashed lines) methods. Ellipses represent tasks and rectangles methods.
We claim that the class of planners we are dealing with share a general, high-level problem-solving method which we call propose-critique-modify (PCM). That is, all planners contain in one way or another these three basic tasks: (i) propose refinement, (ii) critique plan and (iii) modify plan. Planners differ in how they achieve each of these steps; that is, in what PSMs they use to perform these three tasks. These differences also reflect their choices in terms of how planning knowledge is represented. Figure 2 shows a set of PSMs we have found in the literature to realize these tasks.
Figure 2: A task-method decomposition structure for planning.
Actually, the PCM method represents a whole family of methods, in the sense that sometimes not all of its subtasks need to be carried out. For example, the PCM family includes methods that only do propose, or propose and modify. The tasks presented in Figure 2 are described below.
Select goal This task selects a goal from the set of goals to be accomplished. The goal can be either a goal state to be achieved or a goal action to be accomplished by a number of actions. For select goal, three methods can be used: linear select, random select and smart select. STRIPS chronologically selects the last goal state established by its search strategy, which we call linear select. NONLIN [Tate, 1977] randomly selects any goal that has not yet been accomplished. SNLP [McAllester & Rosenblitt, 1991] randomly selects goals from the set of pending goals. It is also possible to select a goal in a more intelligent manner which minimizes possible future modifications ( smart select) [Drummond & Currie, 1989]. In this case, the method uses the static role plan assessment knowledge to select a goal. An example of smart select was proposed by [Kambhampati, 1995], where a goal-selection strategy is described called MTC-based goal selection, which selects a goal state only when it is not necessarily true according to the modal truth criterion.
Propose expansion This task takes the selected goal, and proposes a way to accomplish it using state changes. This can be a new plan step, or an action decomposition which will be added to the plan at the place of the goal action. The propose expansion task can be realized by three alternative methods. Decomposition propose proposes a goal decomposition, which means to propose a more detailed way to accomplish the goal. This method is only applied if the goal is a goal action. Goal-achievement propose, selects an operator whose effect includes the selected goal, constraining the place of the operator in the plan to be necessarily before the selected goal. When a new operator is added to the plan, its preconditions are added to the set of goals. When there is already a step in the plan that achieves the goal, only the ordering constraint is added to the plan. This method is used by SNLP and REFINEMENT SEARCH [Kambhampati, 1995] planning algorithms. The smart propose method is applied by STRIPS and PRODIGY [Carbonell & the PRODIGY Research Group, 1992] which are implemented by means-end analysis [Newell & Simon, 1963], and exploits the difference between the goal state and the current state to select an operator. Within each of the three methods, there are still different ways to propose an expansion, depending on the backtrack points left (e.g., on the decomposition choice, on the operator, or on variable bindings). In the case of adding a step, the different insertion points in the plan can be considered as another backtrack point (e.g., TOCL [Barret & Weld, 1994]).
Test for unachieved goals This task checks the current plan for unachieved goals, and records them in the dynamic role goal. It also tests whether the preconditions (sub-goals) of an operator are already achieved in the current state (when the plan composition is total-order). We identified three methods to realize this task: the MTC-based goal-test, the current-state goal-test and the agenda-based goal test. Planners that exploit causal-links use the simple agenda-based method, because they only need to check for the existence of goals not yet processed; goals, once achieved, are preserved through the causal links.
Interaction critique When checking for conflicts, this task verifies whether the proposed action for accomplishing the goal would interact with other goals in the plan (e.g., one action might undo the precondition of another action). Note that this task involves explicit reasoning about interactions. For realizing the interaction critique task, we identified two methods: (i) the causal-link-based critique, which checks if the proposed plan-step threats any existent causal link; (ii) and the MTC-based critique, which uses the modal truth criterion to check the existence of a step that possibly deletes any achieved goal.
Consistency critique This task checks the consistency of the overall constraints on the plan generated so far. This task differs from the interaction critique in the sense that it can find more general conflicts between the constraints than the deleted-condition conflict. More complex planning systems can also check the consistency of the assigned resources and agents. We identified the constraint propagation method to realize this task.
The backtrack method modifies a plan by backtracking upon failure. Backtracking is possible if there are alternative decision points. It is important to note that a planner can have different possible points to backtrack on, such as the selected goal, the proposed expansion, or the variable bindings.
In the previous section, we have presented a knowledge-level analysis of planning problem-solving methods and we mentioned how their static roles are filled by different kinds of domain knowledge. We can consider the result of the analysis as a library with reusable components, that encompasses the planners based on which it is developed as well as new planners. As pointed out in the introduction, we aim primarily at supporting knowledge engineers in their task to develop knowledge-based systems. Thus, we take an intermediate position between planning experts and domain experts. One should bear in mind that this has consequences for the goals of our analysis and the terminology used. For instance, planning experts would be more interested in support related to the efficient implementation of domain concepts and procedures. However, as a side effect, domain experts and planning experts may also benefit from our analysis of planning.
In previous work [Benjamins, 1994,Benjamins, 1995], we have developed a knowledge acquisition tool TINA ( Tool in Acquisition) that is based on the notions of tasks, PSMs, assumptions, domain knowledge and strategies. TINA works by matching assumptions of PSMs with domain knowledge, and is able to give the following kinds of support: (1) Given a particular domain, generate the applicable strategies. (2) In case no applicable strategies can be found, allow relaxation of some assumptions and generate the legal strategies along with their not fulfilled assumptions. (3) Given a particular strategy, obtain the corresponding assumptions. (4) Given a partial strategy, obtain its possible completions along with the assumptions. By strategy we mean a particular combination of PSMs. In this paper, we focus on (1) and (3). In the sequel, we first explicate how our library for planning can be described in terms of TINA, and then, we perform two experiments corresponding to (1) and (3).
In order to work, TINA needs three kinds of knowledge: (a) a generative grammar describing the legal PSM combinations, (b) conditions on the rewrite rules of the grammar, and (c) domain characterizations.
Figure: 3 Some grammar rules representing planning methods. The term at the left side of the arrow in a grammar rule represents a task. The right part of the arrow represents a composite method (ending in -CM). It consists of a reference to the method's assumptions, and a set of subtasks (ending in -) T and/or primitive PSMs (ending in -PM, and appearing in [ ]).
Figure 3 illustrates some of the composite PSMs represented as rewrite rules. TINA uses the Definite Clause Grammar (DCG) of Prolog to represent the task-method decomposition. Because DCG allows for normal Prolog code within the grammar, we can easily include the assumptions. Figure 3 depicts two alternative PSMs for planning, which belong to the PSM-family, and three PSMs for the test for unachieved goals task.
Figure: 4 Some examples of PSM assumptions.
One result of our analysis of planning is the identification of static roles and their relations to domain knowledge (Figure 1). Our analysis has revealed that there are three main static roles which can be fulfilled by different kinds of domain knowledge. The three main static roles are: (1) plan composition, (2) world description and (3) plan-assessment knowledge. Their respective fillings by domain knowledge are: (1) total or partial order, (2) ADL or HTN and (3) causal links or truth criterion. Below we list six examples of meaningful domain types in planning:
Figure 5: Two examples of domains, characterized by static-role fillings.
The aim of this experiment is to see whether the library can be used to generate reasonable planning strategies given a particular domain type. Suppose we take the domain type (1) from Figure 5: ``partialorder_ADL_truthcriterion''. If we give this to TINA, it outputs four different applicable strategies. The fact the several strategies are found is due to the fact that tasks can have several alternative methods to realize them. One of the strategies generated, is shown below as a tree of methods:
propose_critique_modify_CM propose_I_CM random_select_goal_PM goal_achievement_propose_expansion_PM mtc_based_test_for_unachieved_goals_PM critique_II_CM mtc_based_interaction_critique_PM mtc_based_modify_plan_PM
We can see that this strategy uses the PCM method, consisting of its three subtasks. For each of the subtasks, the possible methods have been checked for applicability by evaluating their assumptions. The generation process has bottomed out in the primitive methods (ending in ` PM'). Actually, the presented strategy is that of TWEAK [Chapman, 1987]. TWEAK randomly selects a goal to be achieved and proposes an expansion that has the goal condition as an effect. It uses MTC-based methods to check for unachieved goals and interactions, and to modify the plan. The critique-II belongs to the family of the critique-I methods, the first being a degenerated version (no consistency critique) of the second.
If we input domain type (2) (Figure 5: ``totalorder_ADL_causallinks''), we are presented with 8 different strategies, among which are those of PRODIGY and STRIPS. Both strategies are part of the PCM family, but they do not contain a critique step.
PRODIGY's strategy is as follows:
propose_modify_CM propose_I_CM random_select_goal_PM smart_propose_expansion_PM current_state_test_for_unachieved_goals_PM backtrack_modify_plan_PM
PRODIGY randomly selects a goal that has not yet been achieved, leaving a backtrack point on the other goals. It uses the smart-propose expansion method (implemented by means-end-analysis [Newell & Simon, 1963]). Unachieved goals are found by checking whether all the goal states are logically implied by the current state. If PRODIGY fails to generate a plan, it backtracks to another goal and continues from there. Note that the fact that the domain knowledge contains plan assessment knowledge in the form of causal links, does not mean that this knowledge necessarily has to be used.
STRIPS' strategy is as follows:
propose_modify_CM propose_I_CM linear_select_goal_PM smart_propose_expansion_PM current_state_test_for_unachieved_goals_PM backtrack_modify_plan_PM
STRIPS uses the linear select method to select a goal to plan for, which does not leave a backtrack point. For the rest, it uses the same methods as PRODIGY. However, the backtrack modify method cannot pursue alternative goals to plan for, but only other operators and variable bindings.
In summary, the main difference between STRIPS and PRODIGY resides in how they backtrack. STRIPS backtracks chronologically on the proposed operator/action and variable bindings. It cannot backtrack on ``select goal'' since it uses the linear select goal method which leaves no alternative choices (this accounts for the `linear' [Tate et al., 1990] characteristic of STRIPS). PRODIGY uses the random select goal method which leaves choice points. Backtracking can thus be performed on the goals selected (which makes PRODIGY a `nonlinear' planner).
Looking at the result, we can say that our library for planning can be used to suggest various specifications of planners, given specific domains. Moreover, the fact that TINA generated several strategies (respectively 4 and 8 for the domain types (1) and (2) of Figure 5), is an indication that the library we are proposing has the competence to explore new, yet unknown, planning strategies.
In this experiment, we use TINA the other way around. We input a certain strategy, and as output we get the assumptions that have to be true in order for the planner to be applicable. Currently we have represented five well-known planners: STRIPS, PRODIGY, SNLP, NONLIN and TWEAK. In this experiment, we will input SNLP and NONLIN.
SNLP is represented as follows:
propose_critique_modify_CM propose_I_CM random_select_goal_PM goal_achievement_propose_expansion_PM agenda_based_test_for_unachieved_goals_PM critique_I_CM constraint_propagation_consistency_critique_PM causal_link_based_interaction_critique_PM causal_link_based_modify_plan_PM
For each constituting method of the strategy, the assumptions are given. Some methods do, however, not have assumptions (e.g., random select goal).
propose_critique_modify_CM: plan_composition=partial_order goal_achievement_propose_expansion_PM: world_description=ADL constraint_propagation_consistency_critique_PM: plan_composition=partial_order causal_link_based_interaction_critique_PM: plan_assessment_knowledge=causal_links causal_link_based_modify_plan_PM: plan_assessment_knowledge=causal_links, plan_composition=partial_order
NONLIN is represented as:
propose_critique_modify_CM propose_I_CM random_select_goal_PM decomposition_propose_expansion_PM mtc_based_test_for_unachieved_goals_PM critique_I_CM constraint_propagation_consistency_critique_PM causal_link_based_interaction_critique_PM causal_link_based_modify_plan_PM
The generated assumptions are:
propose_critique_modify_CM: plan_composition=partial_order decomposition_propose_expansion_PM: world_description=HTN mtc_based_test_for_unachieved_goals_PM: plan_composition=partial_order, plan_assessment_knowledge=truth_criterion constraint_propagation_consistency_critique_PM: plan_composition=partial_order causal_link_based_interaction_critique_PM: plan_assessment_knowledge=causal_links causal_link_based_modify_plan_PM: plan_assessment_knowledge=causal_links, plan_composition=partial_order
In NONLIN, the truth-criterion is used to test for unachieved goals. Critiquing the plan is performed by means of causal links, which, in NONLIN are represented in the so-called ``goal structure'' (GHOST).
In summary, this second experiment demonstrates that it is possible to obtain the assumptions that underly a particular planning strategy. It is then up to the knowledge engineer to decide whether a specific application domain satisfies the assumptions.
The assumptions we discussed so far relate to the problem-solving methods in the sense that they specify what the methods need in order to operate. In order to know whether a method can be used in a given application domain, the domain must guarantee that these assumptions are fulfilled. The method assumptions refer to the domain in three main ways:
The first corresponds to availability epistemological assumptions, whereas the second and third correspond to property epistemological assumptions [Benjamins & Pierret-Golbreich, 1996].
There are several important issues with respect to these assumptions in using them to select a PSM or in characterizing domains. First, it must be kept in mind that in some cases they are `soft' requirements. If one has an implemented planner that would be easier to use for whatever reason, the acquisition process becomes biased towards forcing the domain to fulfill the assumptions required by this planner. Domains can be, to some extent, malleable enough to be bent to these needs. For instance, in some cases it does not matter whether to model the plan sets as add-delete lists or as tasks organized in a hierarchical manner. For instance, Unix can be seen as a planning application domain in which the goals are sets of information of states of the files and directories, and the steps consist of issuing Unix shell commands. Since there seems to be very little hierarchy in the command set, it seem more natural at a first glance to model the Unix commands as operators. On the other hand, as every experienced Unix user knows, one can specify aliases and shell scripts that in practice implement higher level commands for specialized tasks --- the existence of languages such as awk relies to a large extent on this. So the question sometimes is not whether you can model the domain in a certain way, but whether it seems natural or adequate to do so.
Second, and more importantly, availability is not only a matter of looking at what your method needs, but at what is already available in the domain. One should try to use the maximum of the knowledge available in the domain. If (hypothetically speaking) all we have is the Unix manual, flat operators is the best we can come up with. However, if we do have experts that can devise or have already devised efficient shell scripts, it is best to use them. There is a natural trade-off between knowledge and search. In figurative terms, knowledge is the result of previous search recorded in a compiled form. The more (heuristic) knowledge you have, the less you need to rely on search. Of course, a planner can come up alone with the shell scripts, and many learning architectures rely precisely on that. However, if you have them available, why not using them and therefore `save' in search? In this sense, different assumptions and requirements made by different methods present opportunities to make the best use of the knowledge naturally available in the domain.
Third, it is possible (at least in principle) to characterize planning domains using the features, in the same way that we use them to characterize types of methods --- e.g. partial-order versus total-order planners. However, this requires establishing a more clear relation between real domain characteristics and these features oriented to domain representations. For instance, while a method assumption for ``decomposition propose expansion'' requires that knowledge is available and appropriate to fill the world description with an HTN domain model, it is not totally clear what characteristics or features of the application domain indicate whether such assumption will hold. One indication is the way goals are seen: when goals are naturally seen as `actions', HTNs are adequate, whereas when goals are `states', ADLs are more suitable. In other words, we are interested in finding the relations between practical domain characteristics and domain representations, which then are, through our library, related to PSMs for planning.
As we already pointed out above, however, one of the problems is that it is very difficult to characterize domains in hard terms. Further, very little or no research has been spent on characterizing application domains. Therefore, instead of proposing a complete set of domain features that can be used on planning, in the following we will give a first step in this direction by defining a number of domain features that have emerged as important in the course of the research presented in this paper. They will be presented as patterns that connect characteristics of domains to types of planners, passing through the static roles which are relevant and the characteristics of domain models that they invoke. Table 1 gives an overview of these patterns.
Table 1: The relation between application-domain features, static-roles, domain model characteristics (fillings) and methods for planning.
Domains where goals can be expressed either as actions or states, lead to HTN-like planners such as UMCP [Erol et al., 1994]. An example of a type of domain with this characteristic is air campaign planning. Air campaign plans (as studied under the DARPA-Rome Planning Initiative [Tate, 1996]) have their goals typically formulated as actions. Partly as a consequence, such plans are formulated in a hierarchical fashion, in which more specific actions are devised to execute the action-goal, up to the level of primitive actions.
Domains where goals are only expressed in terms of states, allow for STRIPS-like planners such as PRODIGY [Carbonell & the PRODIGY Research Group, 1992] or SNLP [McAllester & Rosenblitt, 1991]. An example of a type of domain with this characteristic is information gathering. In this type of domains (of which the Web is an increasingly important and well-known instance), the goal is to have a specified information with certain constraints, expressed in a query. The planning problem is to devise a series of information retrieval and manipulation steps to obtain the desired information (for an interesting discussion on the challenges of this type of domain, see [Knoblock, 1996]).
This raises the question, why to use STRIPS-like planners at all, since HTN planners can work with both state and action goals. The reason is efficiency. HTN planners are rather complex and heavy systems compared to STRIPS-like systems. So, whenever a planning problem can be solved by a STRIPS-like planner, it seems wise to use that option.
We can also see the use of causal links as an important factor related to memory management of a planner. Causal links are some kind of bookkeeping mechanism that keeps track of which action contributed to specific situations (cf. truth maintenance systems [Doyle, 1979]). Typically, such mechanisms require a lot of static memory. On the other hand, MTC-like planners do not keep track of such dependencies, but generate them on the fly when needed. Consequently, they need few static memory, but much dynamic.
In this paper, we have presented a knowledge-level framework for representing and analyzing planning methods, which can be considered as a library with three types of reusable components: (1) problem-solving methods, (2) their assumptions and (3) domain knowledge. Assumptions form the connection between reasoning knowledge (PSMs) and domain knowledge, and they form an important factor to provide concrete knowledge acquisition support for engineering a planning application.
We have demonstrated, by means of the tool TINA, that the library is able to concretely support the knowledge acquisition process. One kind of support concerned the configuration of an applicable planner for a particular application domain using the PSMs from the library. The application domain in question is matched to the assumptions of PSMs until an adequate configuration of PSMs is found to solve the specific problem. In fact, all possible PSM configurations are found that match the domain at hand, yielding, in several cases, `new' planners. The other kind of knowledge acquisition support concerned the generation of the domain requirements, given a particular planner. The planner is decomposed into its constituting PSMs, and the assumptions of each of the PSMs refer to domain requirements.
The library proposed in this paper has been constructed by abstracting problem-solving strategies (i.e., PSMs) from planning systems and algorithms described in the literature. This characterizes a system-driven approach to the construction of libraries of PSMs. Nevertheless, this is not the only possible approach for building such libraries. For example, one may use a domain-driven approach to obtain PSMs by extracting reasoning knowledge from a domain expert through knowledge acquisition techniques. One may also integrate these two approaches by starting with a domain-driven approach and then comparing the resulting model with a selected system-driven planning model (e.g., one of the models presented in this paper). An approach along these lines has been proposed in [Cottam & Shadbolt, 1996]. It is important to notice, however, that the library presented here can always be extended with new PSMs generated by any of those approaches. Of course, for each PSM added to the library, the respective domain requirements (assumptions) for its selection must also be included.
The library proposed in this paper is basically a modeling aid. That is, it helps a knowledge engineer in selecting a problem-solving method adequate to the application. However, a question that remains is how to operationalize this PSM. The type of library we constructed follows the CommonKADS tradition of separating analysis and design. The conceptual model is supposed to be a product by itself, based upon which the engineer proceeds to design an application and then program a computer system. Design issues are left to a separate model, the design model, which is distinct from the model of expertise where PSMs are specified.
Other research groups such as [Klinker et al., 1991,Swartout & Gil, 1995,Chandrasekaran, 1986] have proposed libraries that use a different approach. Whenever a model has been specified, a number of program modules or routines are provided (or must be programmed) that can immediately operationalize the model into a working system. For example, in the EXPECT architecture [Swartout & Gil, 1995], a method is expressed in terms of goal decompositions up to the point where a level of `primitive' methods, implemented as CommonLisp functions, is reached. By using this strategy, EXPECT can compile a method specified in abstract, knowledge-level, terms into a CommonLisp program.
Of course, each of those approaches has its own advantages and disadvantages. For example, the separation between analysis and design used in CommonKADS does prevent the knowledge engineer from mixing up the analysis with design issues. Moreover, the separation does not restrict the knowledge engineer to use a fixed programming language, operational system or graphics toolkit that may be inappropriate in a given application's setting. On the other hand, authors like [Klinker et al., 1991] have argued that conceptual models hide too many details that have to be decided at design time, and that these design decisions may render the conceptual model inadequate later on.
In order to be able to test the proposed library in a more operational setting, we intend to use the second approach. Specifically, we intend to use the EXPECT system to bridge the gap between the conceptual specification and an implemented system. This will be done in the following steps: (1) construct a set of EXPECT primitive methods corresponding to the lower-level methods that are not decomposed in our library, (2) specify the domain models and domain features used in our library in terms of the domain modeling language used in EXPECT --- namely Loom [MacGregor, 1991], and (3) translate our function structures and control structures into EXPECT methods.
Since EXPECT currently does not have facilities for expressing and using libraries of PSMs such as the one proposed in this paper, the configuration process of creating an EXPECT method will have to be done externally (for example by using TINA) and manually. In a later phase, additional facilities will be built into the EXPECT system to manage libraries of methods, in such a way that the system will be able to handle at least part of the configuration process.
Other further work relates to various extensions of the library. First of all, we will extend the library with more planning methods in order to cover more planners, including skeletal planners, case-based planners, planners that use operators with uncertainty and planners that deal with resources. Secondly, as we include more planning methods, we will also have to include their respective assumptions describing the domain features required by the methods. Thirdly, we will have to deepen the notion of assumption of our planning methods. In this paper, assumptions were described as constraints between static roles and their domain fillings or domain model characteristics (e.g., world_description=HTN). However, we still need to establish a relation between these domain-model characteristics and the application-domain features. We need to find answers to questions such as ``is the HTN world description a domain feature (available or not in the domain) or is it a choice of domain model (that can always be used to structure the domain)?'' We made a first step in this direction, and in future research we will investigate the possibility to express assumptions in terms of such application-domain features.
Richard Benjamins was supported by the Netherlands Computer Science Research Foundation with financial support from the Netherlands Organization for Scientific Research (NWO). Leliane Nunes de Barros acknowledges the CNPq, a Brazilian research foundation. André Valente gratefully acknowledges the support of DARPA through contract DABT63-95-C-0059, as part of the DARPA/Rome Laboratory Planning Initiative (ARPI).
Constructing Planners Through Problem-Solving Methods
This document was generated using the LaTeX2HTML translator Version 95.1 (Fri Jan 20 1995) Copyright © 1993, 1994, Nikos Drakos, Computer Based Learning Unit, University of Leeds.
The command line arguments were:
latex2html -split 0 -show_section_numbers doc.tex.
The translation was initiated by Richard Benjamins on Thu Sep 19 14:54:03 MET DST 1996