## Modeling Intelligent Control of Distributed Cooperative Inferencing (1997)

Citations: | 5 - 0 self |

### BibTeX

@TECHREPORT{Williams97modelingintelligent,

author = {Edward Michael Williams and Edward Michael Williams and Major Usaf and Major Usaf and Robert A. Calico},

title = {Modeling Intelligent Control of Distributed Cooperative Inferencing},

institution = {},

year = {1997}

}

### OpenURL

### Abstract

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x I. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-1 1.1 Goals and Scope . . . . . . . . . . . . . . . . . . . . . . . . . 1-2 1.2 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-3 II. Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-1 2.1 Anytime algorithms . . . . . . . . . . . . . . . . . . . . . . . 2-1 2.2 Algorithm Combinations . . . . . . . . . . . . . . . . . . . . 2-3 2.3 Control Theory . . . . . . . . . . . . . . . . . . . . . . . . . . 2-4 2.4 Intelligent Control . . . . . . . . . . . . . . . . . . . . . . . . 2-5 2.5 Bayesian Networks . . . . . . . . . . . . . . . . . . . . . . . . 2-8 2.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-10 III. Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-1 3.1 Phase 1: Architecture Development . . . . . . . . . . . . . . 3-1 ...

### Citations

7122 |
Probabilistic reasoning in intelligent systems : networks of plausible inference. The Morgan Kaufmann series in representation and reasoning
- Pearl
- 1988
(Show Context)
Citation Context ...ining the appropriate actions. This is definitely in contrast to the traditional expert system mentality which operates in a static, non-real-time environment. 2.5 Bayesian Networks Bayesian Networks =-=(26)-=- are a compact way of representing probabilistic uncertainty in automated reasoning. Such networks consist of directed acyclic graphs of nodes, each representing a random variable (RV) with a finite d... |

1987 |
Genetic Algorithms + Data Structures = Evolution Programs
- Michalewicz
- 1996
(Show Context)
Citation Context ...Bayesian Networks. For this problem domain, the joint probability of the solution is used as the solution quality. The algorithms considered here are an A search(3, 4), a simple Genetic Algorithm (GA)=-=(24)-=-, a hybrid stochastic simulation with local optimization and a Barrier Computation algorithm. All of these algorithms are both anytime and anywhere except for the Barrier algorithm: it is included to ... |

662 | Macready No Free Lunch Theorems for Optimization In
- Wolpert, William
- 1997
(Show Context)
Citation Context ...lt. This leaves a difficult task (computationally) without any general-purpose algorithm to efficiently solve any problem instance (as discussed by Wolpert and Macready in their No Free Lunch Theorems=-=(38)-=-). However, algorithms are available to solve certain specific subclasses of these problems very efficiently. In order to make uncertain inference computationally feasible it is necessary to maximize ... |

500 |
Intractability and Time-Dependent Planning
- Dean
- 1987
(Show Context)
Citation Context ...h. In addition, we also present a brief overview of the target problem domain: Uncertain Reasoning using Bayesian Networks. 2.1 Anytime algorithms Anytime algorithms were first used by Dean and Boddy =-=(2, 7)-=- to provide a means to balance execution time with solution quality in their work on time-dependent planning. In general, anytime algorithms are useful when problems are computationally intractable; t... |

309 |
Planning and Control
- Dean, Wellmann
- 1991
(Show Context)
Citation Context ...ssity for controller design, but provides a level of assurance that the best decisions are being made. Another perspective on control that is relevant to this research is presented by Dean and Wellman=-=(8)-=- and Hayes-Roth(14, 15, 16, 17): control is simply determining a sequence of actions to take to cause the plant to remain in a desired state; which is the definition of planning. The controller uses t... |

260 |
A blackboard architecture for control
- Hayes-Roth
- 1985
(Show Context)
Citation Context ...er design, but provides a level of assurance that the best decisions are being made. Another perspective on control that is relevant to this research is presented by Dean and Wellman(8) and Hayes-Roth=-=(14, 15, 16, 17)-=-: control is simply determining a sequence of actions to take to cause the plant to remain in a desired state; which is the definition of planning. The controller uses the plant model to predict what ... |

255 | Approximating probabilistic inference in bayesian belief networks is np-hard
- Dagum, Ludy
- 1993
(Show Context)
Citation Context ...viding another method for improving the selection accuracy. x Modeling Intelligent Control of Distributed Cooperative Inferencing I. Introduction Uncertain inference using Bayesian Networks is NP-hard=-=(5, 6, 31)-=-; however, it does not change the fact that Bayesian Networks are a common way to represent uncertain knowledge. Example uses include the commercial system Intellipath (formerly called Pathfinder (18)... |

158 | PVM 3 user's guide and reference manual
- Geist, Beguelin, et al.
- 1993
(Show Context)
Citation Context ...ork of Sun and Linux workstations. The first two requirements from the list at the beginning of this chapter were satisfied through the use of the Parallel Virtual Machine (PVM) communications library=-=(10)-=-. The PVM library provides the capability of remote spawning and killing of tasks, as well as a set of communications primitives to allow inter-task communications. Inherent in PVM's design is the abi... |

135 | An Architecture for Adaptive Intelligent Systems
- Hayes-Roth
- 1995
(Show Context)
Citation Context ...er design, but provides a level of assurance that the best decisions are being made. Another perspective on control that is relevant to this research is presented by Dean and Wellman(8) and Hayes-Roth=-=(14, 15, 16, 17)-=-: control is simply determining a sequence of actions to take to cause the plant to remain in a desired state; which is the definition of planning. The controller uses the plant model to predict what ... |

113 | Optimal composition of real-time systems
- Zilberstein, Russell
- 1996
(Show Context)
Citation Context ...thm is not sufficient to solve the problem, multiple algorithms would need to be used; Zilberstein and Russell have researched the use of anytime algorithms as sequential components of a larger system=-=(42)-=-. This introduces the problem of time allocation between the components; how much time should be allocated to the earlier components in order to optimize the results of the later components. This proc... |

112 |
An economics approach to hard computational problems
- Huberman, Lukose, et al.
- 1997
(Show Context)
Citation Context ...ilize the information imported from the other algorithm(s)?(19) ffl When to share information: When is sharing the information beneficial, or is it better to just let the algorithms run independently?=-=(21)-=- The concept underlying information sharing between algorithms is referred to as the anywheresproperty(30). This property refers to an algorithm's ability to accept complete solutions generated elsewh... |

96 |
Probabilistic Inference Using Belief Networks is NP-Hard
- Cooper
- 1987
(Show Context)
Citation Context ...viding another method for improving the selection accuracy. x Modeling Intelligent Control of Distributed Cooperative Inferencing I. Introduction Uncertain inference using Bayesian Networks is NP-hard=-=(5, 6, 31)-=-; however, it does not change the fact that Bayesian Networks are a common way to represent uncertain knowledge. Example uses include the commercial system Intellipath (formerly called Pathfinder (18)... |

92 |
Finding MAPs for belief networks is NPhard
- Shimony
- 1994
(Show Context)
Citation Context ...viding another method for improving the selection accuracy. x Modeling Intelligent Control of Distributed Cooperative Inferencing I. Introduction Uncertain inference using Bayesian Networks is NP-hard=-=(5, 6, 31)-=-; however, it does not change the fact that Bayesian Networks are a common way to represent uncertain knowledge. Example uses include the commercial system Intellipath (formerly called Pathfinder (18)... |

55 | An Executable Model of a Simple Genetic Algorithm. Foundations of Genetic Algorithms -2
- Whitley
- 1993
(Show Context)
Citation Context ...s an old, established research field; but the primary goal of traditional analysis is to determine algorithm complexity(39, 40) or to predict the algorithm's performance over the entire execution time=-=(34, 36)-=-. This is not what is needed for algorithm selection and control; therefore a different approach to algorithm analysis is indicated. We have two motivations for this analysis: algorithm selection and ... |

42 |
Solving the Really Hard Problems with Cooperative Search
- Hogg, Williams
- 1993
(Show Context)
Citation Context ...uses different methods to find the optimal solution to a given problem, much as individuals in a group of people take different approaches when given the same problem. We have observed, as have others=-=(19, 20)-=- that cooperation between algorithms can likewise result in a more rapid solution. The current issues with this approach are: ffl What to share between algorithms: How much and what kind of informatio... |

41 |
Anytime problem solving using dynamic programming
- Boddy
- 1991
(Show Context)
Citation Context ...h. In addition, we also present a brief overview of the target problem domain: Uncertain Reasoning using Bayesian Networks. 2.1 Anytime algorithms Anytime algorithms were first used by Dean and Boddy =-=(2, 7)-=- to provide a means to balance execution time with solution quality in their work on time-dependent planning. In general, anytime algorithms are useful when problems are computationally intractable; t... |

34 | Performance of linear-space search algorithms
- Zhang, Korf
- 1996
(Show Context)
Citation Context ... analytical models of algorithms used on the target problem. Algorithm analysis is an old, established research field; but the primary goal of traditional analysis is to determine algorithm complexity=-=(39, 40)-=- or to predict the algorithm's performance over the entire execution time(34, 36). This is not what is needed for algorithm selection and control; therefore a different approach to algorithm analysis ... |

31 | Intelligent control
- Hayes-Roth
- 1993
(Show Context)
Citation Context ...er design, but provides a level of assurance that the best decisions are being made. Another perspective on control that is relevant to this research is presented by Dean and Wellman(8) and Hayes-Roth=-=(14, 15, 16, 17)-=-: control is simply determining a sequence of actions to take to cause the plant to remain in a desired state; which is the definition of planning. The controller uses the plant model to predict what ... |

29 | Average-case analysis of a search algorithm for estimating prior and posterior probabilities in Bayesian networks with extreme probabilities
- Poole
- 1993
(Show Context)
Citation Context ...evel is reached; but this heuristic is much more sensitive to the features in the conditional probability tables. The determination of how often to advance to a new level is related to the extremeness=-=(27)-=- of the distribution in the conditional probability tables: Case 1 if (P max ) N ! (P min ) N \Gamma1 , the search progresses in a breadth-first fashion since the best entries at each level are not go... |

24 | Better than the best: The power of cooperation
- Hogg, Hubermann
- 1993
(Show Context)
Citation Context ...uses different methods to find the optimal solution to a given problem, much as individuals in a group of people take different approaches when given the same problem. We have observed, as have others=-=(19, 20)-=- that cooperation between algorithms can likewise result in a more rapid solution. The current issues with this approach are: ffl What to share between algorithms: How much and what kind of informatio... |

21 |
A new admissible heuristic for minimal-cost proofs
- Charniak, Husain
- 1991
(Show Context)
Citation Context ...thms for performing belief revision on Bayesian Networks. For this problem domain, the joint probability of the solution is used as the solution quality. The algorithms considered here are an A search=-=(3, 4)-=-, a simple Genetic Algorithm (GA)(24), a hybrid stochastic simulation with local optimization and a Barrier Computation algorithm. All of these algorithms are both anytime and anywhere except for the ... |

21 |
GALGO: a genetic algorithm decision support tool for complex uncertain systems modelled with Bayesian belief networks
- Rojas-Guzman, Kramer
- 1993
(Show Context)
Citation Context ...ore efficient methods must be employed. Current methods for approximating belief revision and updating on Bayesian Networks include best-first search(4), linear programming(29) and genetic algorithms =-=(28, 35)-=-. (Exact algorithms for these problems abound and are too numerous to cite here, see Pearl(26). However, these algorithms are, as a rule, exponential-time algorithms.) The first two methods are determ... |

21 |
Optimality in Parametric Systems
- Vincent, Grantham
- 1981
(Show Context)
Citation Context ...ns, these objectives are typically conflicting. The decision-making process must accommodate this conflict when determining the best algorithm to use. The technique known as multicriteria optimization=-=(33)-=- can be used in this case to determine 6-2 which algorithm combination to use. The optimization is done as a two-step process: first, determine the Pareto-optimal set of configurations; then use a lin... |

19 |
A study of complexity transition on the asymmetric travelling salesman problem
- Zhang, Korf
- 1996
(Show Context)
Citation Context ... analytical models of algorithms used on the target problem. Algorithm analysis is an old, established research field; but the primary goal of traditional analysis is to determine algorithm complexity=-=(39, 40)-=- or to predict the algorithm's performance over the entire execution time(34, 36). This is not what is needed for algorithm selection and control; therefore a different approach to algorithm analysis ... |

17 |
An empirical comparison of three inference methods
- Heckerman
- 1988
(Show Context)
Citation Context ... 31); however, it does not change the fact that Bayesian Networks are a common way to represent uncertain knowledge. Example uses include the commercial system Intellipath (formerly called Pathfinder =-=(18)-=-), a pathological diagnosis system for tissue samples, and power generator real-time monitoring (23, 25). Uncertainty is inevitable in any real-world knowledge base, and it makes the already difficult... |

15 | Opportunistic Control of Action in Intelligent Agents
- Hayes-Roth
- 1994
(Show Context)
Citation Context |

14 | A fast hill-climbing approach without an energy function for finding mpe
- Santos
- 1993
(Show Context)
Citation Context ...obviously unacceptable and more efficient methods must be employed. Current methods for approximating belief revision and updating on Bayesian Networks include best-first search(4), linear programming=-=(29)-=- and genetic algorithms (28, 35). (Exact algorithms for these problems abound and are too numerous to cite here, see Pearl(26). However, these algorithms are, as a rule, exponential-time algorithms.) ... |

14 | Exploiting case-based independence for approximating marginal probabilities
- Shimony, Santos
- 1996
(Show Context)
Citation Context ...in our system uses an implementation of the shared-cost heuristic(3). The heuristic was developed for weighted AND-OR DAGs (WAODAGs); a mapping of the heuristic to Bayesian Networks is straightforward=-=(32)-=-. The next-state generator expands one RV based on a topologically ordered list of RVs. Due to the nature of the heuristic, this ordering is from the leaves of the network to the root. This A algorith... |

13 |
Introduction to Intelligent Control Systems with High Degree of Autonomy,” Introduction to Intelligent and Autonomous Control, P.J.Antsaklis and K.M.Passino
- Antsaklis, Passino
(Show Context)
Citation Context ...nt control requirements for an aircraft when flying at subsonic versus supersonic speeds. In cases like these, an adaptive controller that can accommodate the changes in operating conditions is needed=-=(1)-=-. Key to the operation of most controllers is a model of the plant. This model provides the expected behavior of the plant so that its actual output can be evaluated. The model provides the relationsh... |

8 |
On a distributed anytime architecture for probabilistic reasoning
- Santos, Shimony, et al.
- 1995
(Show Context)
Citation Context ...possibly even the entire network) into a cooperative problem solving environment would create a significant resource for working these difficult problems. This thinking led to the creation of OVERMIND=-=(30)-=-, a distributed architecture for inferencing. The OVERMIND architecture allows the use of different inferencing methods in parallel, controlled by a separate task with the capability to start and stop... |

6 | Design-to-time scheduling and anytime algorithms
- Garvey, Lesser
(Show Context)
Citation Context ... characterizes an algorithm's performance over a set of parameters; they can be generated 2-2 analytically or empirically. Extensive work has been done in empirically characterizing anytime algorithms=-=(9, 11)-=-, but little has been done through algorithm analysis. The logical next step is to use the performance profiles to determine an optimal stopping point. To do this in advance, use the performance profi... |

6 |
Real time estimation of bayesian networks
- Welch
- 1996
(Show Context)
Citation Context ...ore efficient methods must be employed. Current methods for approximating belief revision and updating on Bayesian Networks include best-first search(4), linear programming(29) and genetic algorithms =-=(28, 35)-=-. (Exact algorithms for these problems abound and are too numerous to cite here, see Pearl(26). However, these algorithms are, as a rule, exponential-time algorithms.) The first two methods are determ... |

4 |
Modeling Simple Genetic Algorithms," Foundations of Genetic Algorithms
- Vose
- 1993
(Show Context)
Citation Context ...s an old, established research field; but the primary goal of traditional analysis is to determine algorithm complexity(39, 40) or to predict the algorithm's performance over the entire execution time=-=(34, 36)-=-. This is not what is needed for algorithm selection and control; therefore a different approach to algorithm analysis is indicated. We have two motivations for this analysis: algorithm selection and ... |

3 |
Cost-Based Abduction and
- Charniak, Shimony
- 1994
(Show Context)
Citation Context ...omplex, this method is obviously unacceptable and more efficient methods must be employed. Current methods for approximating belief revision and updating on Bayesian Networks include best-first search=-=(4)-=-, linear programming(29) and genetic algorithms (28, 35). (Exact algorithms for these problems abound and are too numerous to cite here, see Pearl(26). However, these algorithms are, as a rule, expone... |

3 | A.E.Nicholson. Belief Network Inference Algorithms: a Study of Performance Based on Domain Characterisation
- Jitnah
- 1996
(Show Context)
Citation Context ...te the solution. That is, each is potentially exponential, but handles certain classes of networks efficiently. Attempts have been made to empirically characterize the behavior of different algorithms=-=(22)-=-; the results clearly show that the amount of time required to determine the optimal solution is dependent on the characteristics of the network being used. Unfortunately, empirical methods such as th... |

3 |
Experiments with distributed anytime inferencing: Working with cooperative algorithms
- Williams, Santos, et al.
- 1997
(Show Context)
Citation Context ...e optimal answer is found, fewer branches of the search tree would need to be explored to confirm that the optimal answer was found. This analysis explains the mixed results of our initial experiments=-=(37)-=-; all the combinations that showed an increase in performance were combinations containing at least one exact algorithm and at least one approximate algorithm. It was also frequently the case that suc... |

3 |
Composition and monitoring of anytime algorithms
- Zilberstein
- 1995
(Show Context)
Citation Context ...he controller to make a determination of which inferencing methods to use, a model of both the methods' overall performance for a given class of networks and the methods' run-time performance profiles=-=(41)-=- is required. The model allows the controller to choose appropriate methods to begin the inferencing process as well as decide when to change methods in order to converge more rapidly towards the opti... |

2 |
Causal diagrams: clarifying uncertainty
- Kornfeld
- 1991
(Show Context)
Citation Context ...rtain knowledge. Example uses include the commercial system Intellipath (formerly called Pathfinder (18)), a pathological diagnosis system for tissue samples, and power generator real-time monitoring =-=(23, 25)-=-. Uncertainty is inevitable in any real-world knowledge base, and it makes the already difficult task of reasoning over the knowledge base even more difficult. This leaves a difficult task (computatio... |

1 |
and Shlomo Zilberstein. "Anytime Algorithm Development Tools
- Grass
- 1996
(Show Context)
Citation Context ...nally intractable; they provide a means for evaluating the progress of an algorithm during its execution. Anytime algorithms have four characteristics which differentiate them from ordinary algorithms=-=(11)-=-: Quality Measure It is possible to quantify the quality of a solution generated by the anytime algorithm. This could be measure of the closeness of the solution to the optimal answer, or an indicator... |

1 |
A User's Guide to GENESIS." Navy Center for Applied Research
- Grefenstette
- 1987
(Show Context)
Citation Context ...e inevitable error from reality is compensated for by the continual monitoring and control activities. B.2 Genetic Algorithms (GAs) The GAs used in our experiments were simple GAs based on the Genesis=-=(12)-=- framework utilizing both mutation and crossover. We represented a solution (a complete assignment to the underlying Bayesian Network) as an array of integers; each element of the array was converted ... |

1 |
Shlomo Zilberstein. "Monitoring Anytime Algorithms
- Hansen
- 1996
(Show Context)
Citation Context ...ld be estimated and determined in advance, it depends heavily on the accuracy of the estimation methods to get a reasonable result. Another approach is to monitor the progress of the anytime algorithm=-=(13)-=-; this allows us to identify more precisely the point at which the algorithm is not cost-effective. In cases where a single algorithm is not sufficient to solve the problem, multiple algorithms would ... |

1 |
et al. "Commercialization of EPRI's Generator Expert Monitoring System (GEMS)." Expert System Application for the Electric Power Industry
- Morjaia
- 1993
(Show Context)
Citation Context ...rtain knowledge. Example uses include the commercial system Intellipath (formerly called Pathfinder (18)), a pathological diagnosis system for tissue samples, and power generator real-time monitoring =-=(23, 25)-=-. Uncertainty is inevitable in any real-world knowledge base, and it makes the already difficult task of reasoning over the knowledge base even more difficult. This leaves a difficult task (computatio... |