Eugene Fink and Manuela Veloso. Prodigy planning algorithm. Technical Note CMU-CS-94-123, School of Computer Science, Carnegie Mellon University, Pittsburgh, PA 15213, March 1994.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....planning problems, CHICA s search can be controlled by heuristics defining computation, selection, and pruning rules. Comparison of CHICA s search mechanism with other AI planners is difficult because search control is the least documented aspect of AI planners: TWEAK [Chapman 87] and Prodigy [Fink and Veloso 94] are exceptions. TWEAK is a theoretical planning implementation that uses dependency directed breadth first search. Prodigy s planning algorithm uses a combination of backward chaining non linear planning and forward chaining linear planning. Prodigy could accommodate for heuristic rules to ....

Eugene Fink and Manuela Veloso. Prodigy planning algorithm. Technical Note CMU-CS-94-123, School of Computer Science, Carnegie Mellon University, Pittsburgh, PA 15213, March 1994.

....2209 in the case of the problem with 3 discs. Runtime improves from 65 seconds down to 2 seconds. Similarly, when using bf control without ZLIFO the number of explored partial plans reduces from 1554 down to 873. Knoblock (1994) also reports an improvement in performance for the Prodigy planner (Fink Veloso, 1994) when it is using the abstraction hierarchy generated for this domain by the alpine module, which provides in essence the same information as the goal agenda. 9 6. Summary and Comparison to Related Work Many related approaches have been developed to provide a planner with the ability to ....

....1994) concentrate on the positive interactions between operators. The successful matching of e ects to preconditions forms the basis to learn macro operators, see (Dawsson Siklossy, 1977; Korf, 1985) The alpine system (Knoblock, 1994) learns abstraction hierarchies for the Prodigy planner (Fink Veloso, 1994). The approach is based on an ordering of the preconditions and the e ects of each operator, i.e. all e ects of an operator must be in the same abstraction hierarchy and its preconditions must be placed at the same or a lower level than its e ects. This introduces an ordering between the possible ....

Fink, E., & Veloso, M. (1994). Prodigy planning algorithm. Technical report CMU-94-123, Carnegie Mellon University.

....and exploring 2291 partial plans, UCPOP only takes 0.18 0.06 0.01=0.25 s and explores only 48 13 6=67 plans. Unfortunately, any problems or subproblems with more than 3 discs remain beyond the performance of UCPOP. Knoblock [Kno94] also reports an improvement in performance for the Prodigy planner [FV94] when it is using the abstraction hierarchy generated for this domain by the alpine module. 6 Summary and Comparison to Related Work Many related approaches have been developed to provide a planner with the ability to decompose a planning problem by giving it any kind of goal ordering ....

....described in [DS77, Kor85, Kno94] concentrate on the positive interactions between operators. The successful matching of e ects to preconditions forms the basis to learn macro operators in [DS77, Kor85] The alpine system described in [Kno94] learns abstraction hierarchies for the Prodigy planner [FV94]. The approach is based on an ordering of the preconditions and the e ects of each operator, i.e. all e ects of an operator must be in the same abstraction hierarchy and its preconditions must be placed at the same or a lower level than its e ects. This introduces an ordering between the possible ....

E. Fink and M. Veloso. Prodigy planning algorithm. Technical Report CMU94 -123, Carnegie Mellon University, 1994.

....evaluation [5] we were mainly interested in the following questions: 1. Does the extension to a more expressive language lead to a computational overhead 2. How does IP 2 compare to other planners supporting operators with conditional and universally quantified effects such as Prodigy [2] and UCPOP Due to space restrictions we can only sketch a few of the results here and have to refer the reader to [5] To answer Question 1, we compared IP 2 to graphplan on the original graphplan test suite. On most of the examples, IP 2 can outperform graphplan because we have not simply ....

E. Fink and M. Veloso. Prodigy planning algorithm. Technical Report CMU-94123, Carnegie Mellon University, 1994.

....already. In the case of the GPpartial technique, jG 0 n j = jG n j Gamma 1 must hold. The next interesting question is how IP 2 compares to other planners that support universally quantified and conditional effects. We used the Briefcase domain to compare IP 2 to UCPOP [13] Prodigy [5], and graphplan using the equivalent translation of operators into sets. We looked at the simple task of initially having the briefcase at home and several objects in different locations with the goal of finding a roundtrip to bring all objects home. Figure 11 shows the planning task and a plan ....

E. Fink and M. Veloso. Prodigy planning algorithm. Technical Report CMU-94-123, Carnegie Mellon University, 1994.

....Planning backward from the goals has the advantage of producing lower branching factors because there are usually fewer actions applicable to satisfy a goal than a state. Although both the approaches of forward planning and bi directional planning have been used successfully by some planners (Fink Veloso 1994; Blum Furst 1997) the backward planning approach has been the most popular. For HTN planning, the backward planning approach does not have an obvious advantage since the planner constructs plans by decomposing tasks into subtasks by applying decomposition methods. On the other hand, the ....

Fink, E., and Veloso, M. 1994. Prodigy planning algorithm.

....traverse the whole search space and prove that no plan exists that can make both subgoals true simultaneously. The planner has no explicit knowledge of the basic physical law that no object can be in two different locations at the same time. We also compared IP 2 to UCPOP and Prodigy [4] in an extended version of the scheduling domain, originally developed for UCPOP, see Figure 7. Problem time steps actions UCPOP Prodigy IP 2 Graph IP 2 Plan IP 2 Total sched1 3 6 0.94 2.07 0.72 0.02 0.74 sched2 5 8 4.20 10.27 0.75 0.02 0.47 sched3 6 9 4.80 56.71 0.76 0.17 0.93 sched4 7 11 ....

E. Fink and M. Veloso. Prodigy planning algorithm. Technical Report CMU94 -123, Carnegie Mellon University, 1994.

.... the key idea behind the General Problem Solver (GPS) Newell Simon 1961; Ernst Newell 1969) In the late sixties, Fikes, Nilsson, and Raphael embodied the idea in their planner, Strips (Fikes Nilsson 1971) It is still an important technique today, especially as embodied the Prodigy planner (Fink Veloso 1994)) As used by planners, means ends analysis can be described thus: We are given a set of action specifications, an initial situation, and a goal situation description. The problem is to find a sequence of actions that, if carried out starting in the initial situation, would get to a situation ....

....hand, the recent history of research in automated planning tends to have a depressing surplus of completeness results and shortage of heuristic estimators. If you really want completeness, you could plug the Unpop heuristic estimator into a complete goal directed planning framework such as that of (Fink Veloso 1994). Results and Related Work The program has no trouble with the standard toy problems in the literature, where solutions are plans with about five or six steps. My main test domain has been the Manhattan world. Before I discuss how well the program works, let me pause to note how poorly all ....

[Article contains additional citation context not shown here]

Fink, E., and Veloso, M. 1994. Prodigy planning algorithm. Technical Report 94-123, CMU School of Computer Science.

....one was the one to get. Instead of backtracking when the planner cannot buy batteries, it must now replan to acquire the lost resource, i.e. to get more money. Figure 1 illustrates this particular trade off in a general planning scenario using the representation of the search space in PRODIGY (Fink Veloso 1994; Veloso et al. 1995 in press) The figure clearly illustrates the difference between the simulation of execution at planning time, which allows the planner to backtrack upon its choices, and the real execution of steps which triggers the need for replanning instead of simple backtracking. G b I G ....

....a user can interact with a planning algorithm to select the execution of plan steps. We extended the PRODIGY planning algorithm to incorporate real execution. The PRODIGY algorithm is well suited for interleaving planning and execution because it can reason about a simulated execution sequence (Fink Veloso 1994; Stone, Veloso, Blythe 1994) Thus real execution can proceed from a sequence of plan steps which are available for execution. Using its means ends analysis strategy, early in the planning process, PRODIGY selects operators that reduce the differences between the current state and the goal ....

Fink, E., and Veloso, M. 1994. PRODIGY planning algorithm.

.... NoLimit (Veloso, 1989) and prodigy4.0 (Carbonell, Blythe, Etzioni, Gil, Joseph, Kahn, Knoblock, Minton, P erez, Reilly, Veloso, Wang, 1992) NoLimit and prodigy4.0, as opposed to prodigy2.0, do not require the linearity assumption of goal independence and their search spaces are complete (Fink Veloso, 1994). They also have some control over their commitment choices as opposed to the other earlier total order planners. 4. We found that we needed a new name for our algorithm as flecs represents a significant change in philosophy and implementation from prodigy4.0. Veloso Stone 2. A Top Level View ....

....from prodigy4.0. Veloso Stone 2. A Top Level View of flecs prodigy4.0 and flecs differ most significantly from other state of the art planning systems in that they search for a solution to a planning problem by combining backward chaining (or regression) and simulation of plan execution (Fink Veloso, 1994). While back chaining, they can commit to a total ordering of plan steps so as to make use of a uniquely specified world state. These planners maintain an internal representation of the state and update it by simulating the execution of operators found relevant to the goal by backward chaining. ....

[Article contains additional citation context not shown here]

Fink, E., & Veloso, M. (1994). PRODIGY planning algorithm. Technical report CMU-CS94 -123, School of Computer Science, Carnegie Mellon University.

....time, which allows the planner to backtrack upon its choices, and the real execution of steps which triggers the need for replanning instead of simple backtracking. Figure 1 illustrates this particular trade off in a general planning scenario using the representation of the search space in prodigy [4, 16]. The figure clearly illustrates the difference between the simulation of execution at planning time, which allows the planner to backtrack upon its choices, and the real execution of steps which triggers the need for replanning instead of simple backtracking. As shown in Figure 1(b) early ....

....a user can interact with a planning algorithm to select the execution of plan steps. We extended the prodigy planning algorithm to incorporate real execution. The prodigy algorithm is well suited for interleaving planning and execution because it can reason about a simulated execution sequence [4, 14]. Thus real execution can proceed from a sequence of plan steps which are available for execution. Using its means ends analysis strategy, early in the planning process, prodigy selects operators that reduce the differences between the current state and the goal statement. These plan step choices ....

Eugene Fink and Manuela Veloso. PRODIGY planning algorithm. Technical Report CMU-CS-94-123, School of Computer Science, Carnegie Mellon University, 1994.

....are dependent on the state in which the operator is applied. A class (type) hierarchy organizes the objects of the world. These language constructs are important for representing complex and interesting domains. 2 NoLimit was succeeded by the current planner, prodigy4.0 [Carbonell et al. 1992, Fink and Veloso, 1994] The nonlinear planner follows a means ends analysis backward chaining search procedure reasoning about multiple goals and multiple alternative operators relevant to the goals. This choice of operators amounts to multiple ways of trying to achieve the same goal. Therefore, in addition to ....

....and related work prodigy s problem solving method is a combination of means ends analysis, backward chaining, and state space search. prodigy commits to particular choices of operators, bindings, and step orderings as its search process makes use of a uniquely specified state while planning [ Fink and Veloso, 1994 ] prodigy s learning opportunities are therefore directly related to the choices found by the problem solver in its state space search. It is beyond the scope of this paper to discuss what are the potential advantages or disadvantages of our problem solving search method in particular compared ....

Eugene Fink and Manuela Veloso. PRODIGY planning algorithm. Technical Report CMU-CS-94-123, School of Computer Science, Carnegie Mellon University, 1994.

....combined with real execution, replanning may be needed and new steps must be added to the plan, as shown in (b) Real executed steps are shown in diamonds. I is the initial state and G is the goal statement. Operator b reverses the effects of b. the representation of the search space in prodigy [4, 16]. The figure clearly illustrates the difference between the simulation of execution at planning time, which allows the planner to backtrack upon its choices, and the real execution of steps which triggers the need for replanning instead of simple backtracking. As shown in Figure 1(b) early ....

....a user can interact with a planning algorithm to select the execution of plan steps. We extended the prodigy planning algorithm to incorporate real execution. The prodigy algorithm is well suited for interleaving planning and execution because it can reason about a simulated execution sequence [4, 14]. Thus real execution can proceed from a sequence of plan steps which are available for execution. Using its means ends analysis strategy, early in the planning process, prodigy selects operators that reduce the differences between the current state and the goal statement. These plan step choices ....

Eugene Fink and Manuela Veloso. PRODIGY planning algorithm. Technical Report CMU-CS-94-123, School of Computer Science, Carnegie Mellon University, 1994.

....cases occurs by extending the base level planner with the ability to examine its internal decision cycle, recording the justifications for each decision during its search process. Prodigy Analogy has been re implemented in Prodigy4.0, a state space nonlinear planner (Carbonell et al. 1992; Fink Veloso 1994). Prodigy4.0 s planning reasoning cycle involves several decision points, namely: the goal to select from the set of pending goals; the operator to choose to achieve a particular goal; the bindings to choose in order to instantiate the chosen operator; apply an operator whose preconditions are ....

Fink, E., and Veloso, M. 1994. PRODIGY planning algorithm. Technical Report CMU-CS-94-123, School of Computer Science, Carnegie Mellon University.

....part 1) Figure 8: Instantiating a newly added operator. A summary of the backward chaining algorithm is presented in Table 2. It may be shown that the use of this back chaining algorithm with the execution simulator described in the previous subsection guarantees the completeness of planning (Fink and Veloso, 1994). Notice that the decision point in Step 1 of the algorithm may be required for the efficiency of the depth first planning, but not for completeness. If PRODIGY generates the entire tail before applying it, we do not need branching on Step 1. We may pick preconditions in any order, since it does ....

Fink, E. and Veloso, M. (1994). PRODIGY planning algorithm. Technical Report CMU-CS-94-123, School of Computer Science, Carnegie Mellon University.

....are two objects, ob4 and ob7, one truck tr9, and one airplane pl1. There is one city c3 with a post office p3 and an airport a3. In the initial state, ob4 is at p3 and the goal is to have ob4 inside of tr9. 1 NOLIMIT was succeeded by the current planner, PRODIGY4.0 (Carbonell et al. 1992; Fink Veloso 1994). state (and (at obj ob4 p3) at truck tr9 a3) same city a3 p3) at obj ob7 a3) at airplane pl1 a3) tr9 pl1 tr9 p3 a3 city c3 (inside truck ob4 tr9) goal ob4 ob4 ob7 Figure 2: Example: The goal is to load one object into the truck. Initially the truck is not at the object s location. The ....

....and related work PRODIGY s problem solving method is a combination of means ends analysis, backward chaining, and state space search. PRODIGY commits to particular choices of operators, bindings, and step orderings as its search process makes use of a uniquely specified state while planning (Fink Veloso 1994). PRODIGY s learning opportunities are therefore directly related to the choices found by the problem solver in its state space search. It is beyond the scope of this paper to discuss what are the potential advantages or disadvantages of our problem solving search method in particular compared ....

Fink, E., and Veloso, M. 1994. PRODIGY planning algorithm.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC