Eric A. Hansen and Shlomo Zilberstein. Monitoring and control of anytime algorithms: A dynamic programming approach. Artificial Intelligence, 126(1--2):139--157, 2001.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....actions, it has to select among such actions. Reasoning about which deliberation actions to take is called metareasoning. Decision theory [7, 10] provides a normative basis for metareasoning under uncertainty, and decisiontheoretic deliberation control has been widely studied in AI (e.g. [2, 4 6, 8, 9, 12 15, 18 20] ) However, the approach of using metareasoning to control reasoning is impractical if the metareasoning problem itself is prohibitively complex. While this issue is widely acknowledged (e.g. 8, 12 14] there are few theoretical results on the complexity of metareasoning. We derive ....

....condition the performance prediction on features of the instance and this is necessary if the deliberation control is to be fully normative. Research has already been conducted on conditioning performance profiles on instance features [8, 9, 15] or results of deliberation on the instance so far [4, 8, 9, 15, 18 20] . problem instances to get a total performance of at least K; that is, whether there exists a vector (N 1 ,N 2 , N m)with K. A reasonable approach to representing the performance profiles is to use piecewise linear performance profiles. They can model any performance profile ....

E Hansen and S Zilberstein. Monitoring and control of anytime algorithms: A dynamic programming approach. Artificial Intelligence, 126:139-157, 2001.

....resources for quality of results. The meta level control procedures are called performance profiles, and are models of how the quality of a solution produced by an algorithm improves with computation time. There has been a lot of work on performance profile based control of computation [20, 2, 1, 3]. To represent performance profiles we use a performance profile tree [7] The advantage of this approach is that, unlike earlier approaches, it allows optimal conditioning on solution quality so far, the results of execution so far, as well as conditioning on the problem instance and other ....

E. Hansen and S. Zilberstein. Monitoring and control of anytime algorithms: A dynamic programming approach. Artificial Intelligence, 126:139--157, 2001.

.... an algorithm s future performance can be deterministically predicted using a performance profile [10,11] assuming that an anytime algorithm s future performance does not depend on the run on that instance so far [4,9,38, 39] or that performance is conditioned on quality so far but not the path [8], or resorting to asymptotic notions of bounded optimality [27] While such simplifications can be acceptable in single agent settings as long as the agent performs reasonably well, any deviation from full normativity can be catastrophic in multiagent settings. If the designer cannot guarantee ....

....time. As will be discussed later, each agent uses this information to decide how to allocate its computation at every step of the game, optimally striking a tradeoff between computation time and solution quality. A common representation of performance profiles is a table of discrete values [8,39]. This approach requires discretizing time into a finite number of time steps and solution quality into a finite number of solution levels. For each time step and each level of solution quality, the table contains the probability that the solution will be of that quality. The resolution of the ....

E. Hansen, S. Zilberstein, Monitoring and control of anytime algorithms: A dynamic programming approach, Artificial Intelligence 126 (2001) 139--157.

....to select among such actions. Reasoning about which deliberation actions to take is called metareasoning. Decision theory [7, 10] provides a normative basis for metareasoning under uncertainty, and decision theoretic deliberation control has been widely studied in AI for the last 15 years (e.g. [2, 4 6, 8, 9, 12 15, 19 21]) However, the approach of using metareasoning to control reasoning is impractical if the metareasoning problem itself is prohibitively complex. While this issue is widely acknowledged (e.g. 8, 12 14] there are few theoretical results on the complexity of metareasoning. We derive hardness ....

....condition the performance prediction on features of the instance and this is necessary if the deliberation control is to be fully normative. Research has already been conducted on conditioning performance profiles on instance features [8, 9, 15] or results of deliberation on the instance so far [4, 8, 9, 15, 19 21]. 0 0.5 1 1.5 2 2.5 3 3.5 4 0 1 2 3 4 5 6 Savings Computation Time (hours) instance 1 instance 2 instance 3 Figure 1: Performance profiles for routing problems Then the maximum savings we can obtain with 5 hours of deliberation time is 2.5, for instance by spending 3 hours on ....

Eric Hansen and Shlomo Zilberstein. Monitoring and control of anytime algorithms: A dynamic programming approach. Artificial Intelligence, 126:139--157, 2001.

....profiles that describe how computation changes the valuations. Each agent uses this information to decide how to allocate its computation at every step in the process, based on the results of its computation so far. There has been much work on performance profile based deliberation control [24, 5, 2, 6]. To represent the performance profiles we use a tree structure [11] The advantage of this approach is that it allows optimal conditioning on results of execution so far, and allows conditioning on the actual problem instance. We specify two different types of performance profiles, stochastic ....

....profile tree allows one to capture this uncertainty. The tree can be used to determine # ## # # ### denoting the probability that running the algorithm for # time steps produces a solution of value # # # . Unlike earlier performance profile models for controlling anytime algorithms (e.g. [24, 2, 6, 5]) the performance profile tree supports conditioning on the path of valuation quality so far. The performance profile tree that applies given a path of computation is the subtree rooted at the current node #. This subtree is denoted by # # # ###. If an agent is at a node # with value #, then ....

Eric Hansen and Shlomo Zilberstein. Monitoring and control of anytime algorithms: A dynamic programming approach. Artificial Intelligence, 126:139--157, 2001.

....due to computational limitations. 2 deterministically predicted using a performance profile [10, 11] assuming that an anytime algorithm s future performance does not depend on the run on that instance so far [4, 39, 38, 9] or that performance is conditioned on quality so far but not the path [8], or resorting to asymptotic notions of bounded optimality [27] While such simplifications can be acceptable in single agent settings as long as the agent performs reasonably well, any deviation from full normativity can be catastrophic in multiagent settings. If the designer cannot guarantee ....

....time. As will be discussed later, each agent uses this information to decide how to allocate its computation at every step of the game, optimally striking a tradeoff between computation time and solution quality. A common representation of performance profiles is a table of discrete values [39, 8]. This approach requires discretizing time into a finite number of time steps and solution quality into a finite number of solution levels. For each time step and each level of solution quality, the table contains the probability that the solution will be of that quality. The resolution of the ....

Eric Hansen and Shlomo Zilberstein. Monitoring and control of anytime algorithms: A dynamic programming approach. Artificial Intelligence, 126:139--157, 2001.

....Each agent uses this information to decide how to allocate its computation at every step in the process, based on the results of its computation so far. There has been muchwork on performance pro le based deliberation control [Zilberstein and Russell, 1996; Boddy and Dean, 1994; Horvitz, 1987; Hansen and Zilberstein, 2001] To represent the performance pro les we use a tree structure [Larson and Sandholm, 2000] The advantage of this approach is that it allows optimal conditioning on results of execution so far, and can condition on the actual problem instance. We index the problem by # and # where # is an ....

Eric Hansen and Shlomo Zilberstein. Monitoring and control of anytime algorithms: A dynamic programming approach. Articial Intelligence, 126:139-157, 2001.

....we use in Section 5. Russell and Wefald have developed a general framework for meta reasoning and applied it to control search in game playing programs [17] Hansen and Zilberstein have developed a comprehensive solution to the problem of monitoring and control of interruptible anytime algorithms [8,9]. The solution covers several probabilistic representations of performance profiles and utility functions. An important part of the solution is the ability to factor the cost of monitoring the computation and or the state of the environment. Sandholm and Lesser have presented an algorithm for ....

Eric A. Hansen and Shlomo Zilberstein. Monitoring and control of anytime algorithms: A dynamic programming approach. To appear in Artificial Intelligence, 2001.

No context found.

Eric A. Hansen and Shlomo Zilberstein. Monitoring and control of anytime algorithms: A dynamic programming approach. Artificial Intelligence, 126(1--2):139--157, 2001.

No context found.

Hansen, E.A., Zilberstein, S.: Monitoring and control of anytime algorithms: A dynamic programming approach. Artificial Intelligence 126 (2001) 139--157

No context found.

E. Hansen and S. Zilberstein. Monitoring and control of anytime algorithms: A dynamic programming approach. Artificial Intelligence, 126:139--157, 2001.

No context found.

Eric Hansen and Shlomo Zilberstein. Monitoring and control of anytime algorithms: A dynamic programming approach. Artificial Intelligence, 126:139--157, 2001.

No context found.

Eric Hansen and Shlomo Zilberstein. Monitoring and control of anytime algorithms: A dynamic programming approach. Artificial Intelligence, 126:139--157, 2001.

No context found.

E. Hansen and S. Zilberstein. Monitoring and control of anytime algorithms: A dynamic programming approach. Artificial Intelligence, 126:139--157, 2001.

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