Ronald A. Howard and James E. Matheson, editors. Readings on the Principles and Applications of Decision Analysis. Strategic Decision Group, Menlo Park, CA, 1984.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....F is the utility function for the value node V, defined in terms of its parents. A number of algorithms for evaluating an influence diagram determining the agent s maximum utility, and the corresponding decisions have been proposed. We will not present any here, but direct the interested reader to [59, 94] for examples of such algorithms. 31 2.5 Structured Representations of Actions In Sections 2.1 and 2.2 we have presented a very general model of MDPs. Unfor tunately, this generality has the disadvantage that it is can be quite an inefficient way of representing a problem. This is because ....

Ronald A. Howard and James E. Matheson, editors. Readings on the Principles and Applications of Decision Analysis. Strategic Decision Group, Menlo Park, CA, 1984.

....model on these influence diagrams is reasonable. The estimate of the incremental value of refinement seems to be somewhat optimistic with respect to value. We are pursuing this issue further. 2 BACKGROUND An influence diagram is a DAG representing a sequential decision problem under uncertainty [Howard Matheson, 1984] . An ID models the subjective beliefs, preferences, and available actions from the perspective of a single decision maker. A policy prescribes an action for each possible combination of observation. An optimal policy max imizes the decision maker s expected value, without regard to the cost of ....

....maker s expected value, without regard to the cost of finding such a policy. If computational costs are negligible, the decision maker s expected value depends only on the expected value of an optimal policy. Traditional algorithms which compute the optimal policy using dynamic programming [Howard Matheson, 1984; Shachter, 1986] usually assume computational costs to be negligible. 2.1 FLEXIBLE COMPUTATION In situations in which there is uncertainty about the state of the world and uncertainty about the possible outcomes of action, it has been argued that a rational decision maker should act so as to ....

Howard, R., and Matheson, J., eds. 1984. Readings on the Principles and Applications of Decision Analysis. CA: Strategic Decisions Group.

....would be available by traditional methods. 1 INTRODUCTION The representational tools which decision analysts and AI practitioners have devised can represent large decision problems. When costs of computation are not taken into account, optimal policies can be determined using dynamic programming [Howard Matheson, 1984; Shachter, 1986] When the costs of computation are not negligible, the cost of computing the optimal policy using dynamic programming may be prohibitive. We have developed an algorithm which can be used to compute policies for large multi stage decision problems under uncertainty represented ....

....influence diagrams and the decision tree representation of decision functions. Section 2 presents the random access algorithm. Empirical results are presented in Section 3. 1. 1 INFLUENCE DIAGRAMS An influence diagram (ID) is a DAG representing a sequential decision problem under uncertainty [Howard Matheson, 1984] . An ID models the subjective beliefs, preferences, and available actions from the perspective of a single decision maker. Nodes in an ID are of three types. Random variables, which the decision maker cannot control, are represented by circle shaped chance nodes. Decisions, i.e. sets of ....

Howard, R., and Matheson, J., eds. 1984. Readings on the Principles and Applications of Decision Analysis. CA: Strategic Decisions Group.

.... recognition that probability and decision theory, hitherto applied primarily to problems of statistical estimation [129, 122] also could be applied to real world decision problems [78, 120] Since its inception, decision analysis has grown into an established academic and professional discipline [83, 152, 88]. There are a number of commercial consulting and research firms that perform decision analyses for government and private clients. Some large corporations routinely apply decision analysis to scheduling, capital expansion, and research and development decisions. The emphasis has been on assisting ....

....for propagating evidence and finding optimal decisions. Finally, we review research on explaining the results of decision theoretic inference. 4. 1 Knowledge Representation for Decision Theoretic Problems Howard has called the complete model of a decision problem the decision basis [83]. A comprehensive decision basis consists of components that represent the alternatives, states, preferences, and relationships in a decision situation. Decisions are the alternative courses of action available to the decision maker. The alternative states of the world are those factors or ....

R.A. Howard and J.E. Matheson, editors. Readings on the Principles and Applications of Decision Analysis. Strategic Decisions Group, Menlo Park, Ca., 1984.

....model on these influence diagrams is reasonable. The estimate of the incremental value of refinement seems to be somewhat optimistic with respect to value. We are pursuing this issue further. 2 BACKGROUND An influence diagram is a DAG representing a sequential decision problem under uncertainty [Howard Matheson, 1984] . An ID models the subjective beliefs, preferences, and available actions from the perspective of a single decision maker. A policy prescribes an action for each possible combination of observation. An optimal policy max imizes the decision maker s expected value, without regard to the cost of ....

....maker s expected value, without regard to the cost of finding such a policy. If computational costs are negligible, the decision maker s expected value depends only on the expected value of an optimal policy. Traditional algorithms which compute the optimal policy using dynamic programming [Howard Matheson, 1984; Shachter, 1986] usually assume computational costs to be negligible. 2.1 FLEXIBLE COMPUTATION In situations in which there is uncertainty about the state of the world and uncertainty about the possible outcomes of action, it has been argued that a rational decision maker should act so as to ....

Howard, R., and Matheson, J., eds. 1984. Readings on the Principles and Applications of Decision Analysis. CA: Strategic Decisions Group.

....of the situation, the outcome will be as good as can be expected. 1 This principle is simply stated, although there is much to be said about the application of this principle. A decision problem can be posed as a decision tree [40] a Markov decision process [35] an influence diagram [21], or as a variant of these representations. The textbook approaches which apply the principle of maximum expected utility to problems in these representations typically make the assumption of negligible computational costs [35, 21, 45] There is considerable interest in how the principle of ....

....tree [40] a Markov decision process [35] an influence diagram [21] or as a variant of these representations. The textbook approaches which apply the principle of maximum expected utility to problems in these representations typically make the assumption of negligible computational costs [35, 21, 45]. There is considerable interest in how the principle of maximum expected utility can be realized in practice [11, 41, 42, 18, 47] since the assumption of negligible computational costs is not always appropriate. In domains such as medical informatics, decision problems may be so large that ....

[Article contains additional citation context not shown here]

R.A. Howard and J.E. Matheson, editors. Readings on the Principles and Applications of Decision Analysis. Strategic Decisions Group, CA, 1984.

....would be available by traditional methods. 1 INTRODUCTION The representational tools which decision analysts and AI practitioners have devised can represent large decision problems. When costs of computation are not taken into account, optimal policies can be determined using dynamic programming [Howard Matheson, 1984; Shachter, 1986] When the costs of computation are not negligible, the cost of computing the optimal policy using dynamic programming may be prohibitive. We have developed an algorithm which can be used to compute policies for large multi stage decision problems under uncertainty represented ....

....influence diagrams and the decision tree representation of decision functions. Section 2 presents the random access algorithm. Empirical results are presented in Section 3. 1. 1 INFLUENCE DIAGRAMS An influence diagram (ID) is a DAG representing a sequential decision problem under uncertainty [Howard Matheson, 1984] . An ID models the subjective beliefs, preferences, and available actions from the perspective of a single decision maker. Nodes in an ID are of three types. Random variables, which the decision maker cannot control, are represented by circle shaped chance nodes. Decisions, i.e. sets of ....

Howard, R., and Matheson, J., eds. 1984. Readings on the Principles and Applications of Decision Analysis. CA: Strategic Decisions Group.

....parameters needed to specify the underlying probability model becomes impractical. That is why several approximations propose to factorize the probability distribution according to a probability model. 3. 3 Bayesian networks This section will introduce the probabilistic graphical model paradigm [35, 42, 50] that has been used during the last decade as a popular representation for encoding uncertainty knowledge in expert systems [32] Only probabilistic graphical models whose structural part is a directed acyclic graph will be considered, as these adapt properly to EDAs the only documented EDA that ....

R. Howard and J. Matheson, Readings on the Principles and Applications of Decision Analysis, Vol. II, chapter In uence diagrams, 721-762, Strategic Decision Group, Menlo Park, CA, 1981. R. Howard and J. Matheson eds.

.... various CADrelated topics such as: mapping the enterprise information language EXPRESS (Wermelinger and Bejan, 1993) and modeling the knowledge interchange format KIF (Sowa, 1993) Since conceptual structures cannot represent uncertainties well, others graph based models called influence diagrams (Howard and Matheson, 1983) can be used to represent probabilistic knowledge for probabilistic inference. Influence diagrams are extensions of Bayesian networks and have been used in artificial intelligence, decision analysis and statistics. Figure 6 depicts an approximate hierarchy of some graph based models. General ....

Howard, R. A. and Matheson, J. E., editors (1983). Readings on The Principles and Applications of Decision Analysis, Strategic Decisions Group, Menlo Park, CA.

....in which the robot is wet and the owner has coffee over those where the robot stays dry and its owner is without coffee: thus, delivering coffee is a higher priority objective for the robot than staying dry. The reward node in this example is related to the value nodes of influence diagrams [45, 69]. In influence diagrams, these nodes generally represent (long term) value, whereas we use them to represent immediate reward (note that we assume stationarity of the reward process) In both cases, the independence of reward and certain state variables is exploited. Some work on influence ....

Ronald A. Howard and James E. Matheson, editors. Readings on the Principles and Applications of Decision Analysis. Strategic Decision Group, Menlo Park, CA, 1984.

....tools. Second, there are various kinds of graph representations associated with computer tools that may become readily available for users of graph based computational techniques. Two particular types of graph based models we refer to are: conceptual structures (Sowa, 1984) and influence diagrams (Howard and Matheson, 1983). Conceptual structures are a graphic system for representing concepts and relations that is general as predicate calculus. There have been studies on the use of conceptual graphs for various CAD related topics such as: mapping the enterprise information language EXPRESS (Wermelinger and Bejan, ....

Howard, R. A. and Matheson, J. E., editors (1983). Readings on The Principles and Applications of Decision Analysis, Strategic Decisions Group, Menlo Park, CA.

....blocks as a problem size grows. 1 INTRODUCTION Recently, there has been a growing interest in optimization methods that explicitly model the good solutions found so far and use the constructed model to guide the further search (Baluja, 1994; Harik et al. 1997; Muhlenbein Paa, 1996; Muhlenbein et al. 1998; Pelikan Muhlenbein, 1999) This line of research in stochastic optimization was strongly motivated by results achieved in the field of evolutionary computation. However, the connection between these two areas has sometimes been obscured. Moreover, the capabilities of model building have often ....

....Covering pairwise interactions still does not preserve higher order partial solutions. Moreover, interactions of higher order do not necessarily imply pairwise interactions that can be detected at the level of partial solutions of order two. In the factorized distribution algorithm (FDA) Muhlenbein et al. 1998), a factorization of the distribution is used for generating new solutions. The distribution factorization is a conditional distribution constructed by analyzing the problem decomposition. The FDA is capable of covering the interactions of higher order and combining important partial solutions ....

[Article contains additional citation context not shown here]

In Howard, R. A., & Matheson, J. E. (Eds.), Readings on the Principles and Applications of Decision Analysis, Volume II (pp. 721--762). Menlo Park, CA: Strategic Decisions Group. Kargupta, H. (1998). Revisiting the GEMGA: Scalable evolutonary optimization through linkage learning.

....made. This is often the case in domains such as product configuration or medical diagnosis (to name but two) Extracting preference information from users is generally arduous, and human decision analysts have developed sophisticated techniques to help elicit this information from decision makers [11]. A key goal in the study of computerbased decision support is the construction of tools that allow the preference elicitation process to be automated, either partially or fully. In particular, methods for extracting, representing and reasoning about the preferences of naive users is especially ....

....are important because they should aid in the elicitation process for naive users. 1 See www.activebuyersguide.com. The tools there also ask for some semi quantitative information about preferences. Preference elicitation is a complex task and is a key focus in work on decision analysis [13, 11, 9], especially elicitation involvingexpert users. Automating the process of preference extraction can be very difficult. Straightforward approaches involvingthe direct comparison of all pairs of outcomes are generally infeasible for a number of reasons, including the exponential number of outcomes ....

Ronald A. Howard and James E. Matheson, editors. Readings on the Principles and Applications of Decision Analysis. Strategic Decision Group, Menlo Park, CA, 1984.

....problems in which all chance and decision variables have a discrete state space. For problems in which some of the decision and or chance variables are continuous, several approximate approaches have been proposed. The traditional approach is to discretize the continuous variables to a few states [Howard and Matheson, 1983, 1984, Miller and Rice 1983, Keefer 1994, Smith 1991] A related approach is to summarize continuous distributions by their first few moments, summarize continuous utility functions by their first few derivatives, and then either the moments and derivatives are used directly [Howard 1971] or the ....

Howard, R. A. and J. E. Matheson (1984), Readings on the Principles and Applications of Decision Analysis, 2, Strategic Decisions Group, Menlo Park, CA.

....problems in which all chance and decision variables have a discrete state space. For problems in which some of the decision and or chance variables are continuous, several approximate approaches have been proposed. The traditional approach is to discretize the continuous variables to a few states [Howard and Matheson, 1983, 1984, Miller and Rice 1983, Keefer 1994, Smith 1991] A related approach is to summarize continuous distributions by their first few moments, summarize continuous utility functions by their first few derivatives, and then either the moments and derivatives are used directly [Howard 1971] or the ....

Howard, R. A. and J. E. Matheson (1983), Readings on the Principles and Applications of Decision Analysis, 1, Strategic Decisions Group, Menlo Park, CA.

....the time quality tradeoffs. We briefly sketch ways to compute the error bound associated with a particular abstraction as well as improve this bound through judicious selection of new relevant atoms. Deciding which atoms to add to the set IR is essentially a value of information calculation [27, 41]. We could imagine for example being given a time bound and wanting the best possible solution computable within that time. Since a time bound restricts the number of atoms that can be considered, we want a set of relevant atoms that satisfies the size restriction and has the lowest possible error ....

....Thus we cannot guarantee that determining the most important single atom will aid in constructing the most valuable set of two (or more) atoms. Determining a set of variables of fixed size with greatest information value generally requires exhaustive search through the space of possible sets [27]. If time restrictions require a set IR of size k, we potentially have to enumerate all size k subsets of the set of atoms P and determine their span, choosing a set with smallest span. Of course, atoms not mentioned in the reward discriminant can have no impact on span, so we can restrict ....

[Article contains additional citation context not shown here]

Ronald A. Howard and James E. Matheson, editors. Readings on the Principles and Applications of Decision Analysis. Strategic Decision Group, Menlo Park, CA, 1984.

.... can be combined with utilities, Applications of Bayesian Networks to Health Care 4 NSI TR 1997 02 such as quality adjusted life years, or the cost of a cataract operation, to determine the optimal decision when allocating resources, according to the precepts of decision analysis [Howard 1984]. This cannot be performed with, say, a neural network s score of 15 for cataracts. Figure 2 Ergo s graphical network editor 3. Ergo TM and Cogito TM: Tools for Constructing Bayesian Network s The semantic clarity, relative computational efficiency, and intuitiveness of expert systems based on ....

Howard RA, Matheson, JE. (1984). Readings on the Principles and Applications of Decision Analysis. Menlo Park, CA: Strategic Decisions Group.

....the evaluation methodology for the preference acquisition technique and the decision support system. Section 7 contains some concluding remarks. 2 Background 2. 1 Decision Theory Decision theory is a discipline of study combining ideas from operations research, statistics, and computer science [25,59,63]. It provides an explicit methodology for selecting an optimal action (or set of actions) from 3 many competing actions which have di#erent outcomes. There are four types of decision problems depending on the presence or absence of uncertainty, and on whether the decision problem has a single ....

Howard RA, Matheson JE, eds. Readings on the Principles and Applications of Decision Analysis. Strategic Decisions Group, Menlo Park, CA, 1983.

....as decision making involves not only knowledge of the uncertainties in a problem under study, but also knowledge of the decisions that are at a decision maker s disposal and of the desirability of their uncertain consequences. The framework of influence diagrams is tailored to decision making [Howard and Matheson 1981]. It provides a formalism for capturing the various types of knowledge involved in a decision problem and o#ers algorithms for computing preferred decisions. The framework is closely related to the belief network framework; in fact, influence diagrams may be looked upon as enhanced belief ....

....a probability distribution. It thus provides for computing any probability of interest. To this end e#cient algorithms are available [Pearl 1988, Lauritzen and Spiegelhalter 1988] A Bayesian belief network may be extended to an influence diagram to allow for decision making under uncertainty [Howard and Matheson 1981]. The formalism of influence diagrams provides for encoding not only a probability distribution on a set of variables, but also the decisions that a decision maker can take and the desirability of their uncertain consequences. As a belief network, an influence diagram consists of a qualitative ....

Howard, R.A. and Matheson, J. 1981. Readings on the Principles and Applications of Decision Analysis, Strategic Decisions Group, Menlo Park.

....made. This is often the case in domains such as product configuration or medical diagnosis (to name but two) Extracting preference information from users is generally arduous, and human decision analysts have developed sophisticated techniques to help elicit this information from decision makers [10]. A key goal in the study of computerbased decision support is the construction of tools that allow the preference elicitation process to be automated, either partially or fully. In particular, methods for extracting, representing and reasoning about the preferences of naive users is especially ....

....ask for some semi quantitative information about preferences. they should aid in the elicitation process for naive users, tools for representing and reasoning about ceteris paribus preferences are important. Preference elicitation is a complex task and is a key focus in work on decision analysis [12, 10, 8], especially elicitation involvingexpert users. Automating the process of preference extraction can be very difficult. Straightforward approaches involvingthe direct comparison of all pairs of outcomes are generally infeasible for a number of reasons, including the exponential number of outcomes ....

R. A. Howard and J. E. Matheson, eds. Readings on the Principles and Applications of Decision Analysis. Strategic Decision Group, Menlo Park, CA, 1984.

....to two stages (corresponding to states s and do(a; s) is appropriate given our semantics. We also point out that, as with our description of the situation calculus, such a BN does not uniquely specify a transition matrix for action a (more in the next section) Finally, influence diagrams (IDs) [13, 27] have been used in probabilistic inference and decision analysis to represent decision problems. Similar in structure to BNs, they have additional types of variables, represented by value nodes and decision nodes; we are only interested in decision nodes here. A decision node (or action node) is ....

Ronald A. Howard and James E. Matheson, editors. Readings on the Principles and Applications of Decision Analysis. Strategic Decision Group, Menlo Park, CA, 1984.

....and a set or preferences qualifying the relative goodness of particular outcomes. Extracting these four types of information can be extremely difficult and time consuming, and human decision analysts have developed sophisticated techniques to help elicit this information from decision makers [8]. For many application domains, however, once a model of the system is known, it is unlikely to change in substantial ways. This can provide considerable leverage in the construction of automated decision making agents (DAs) for a given application; the models can be hard wired into the DA ....

....to extract user preferences on a case by case basis. This type of information cannot be hard wired or precompiled, since the aim is to act on the behalf of, or advise, a particular user. Preference elicitation is a very difficult task in general and is a key focus in work on decision analysis [10, 8, 6]. Automating the process of preference extraction can be very difficult. Straightforward approaches involving the direct comparison of all pairs of outcomes are generally infeasible for a number of reasons, including the exponential number of outcomes (in the number of relevant variables or ....

R. A. Howard and J. E. Matheson, eds. Readings on the Principles and Applications of Decision Analysis. Strategic Decision Group, Menlo Park, CA, 1984.

....et al. 1976, Conners and Harlow, 1980, du Buf et al. 1990, Ohanian and Dubes, 1992) Using the co occurrence features, an accuracy of 72.5 was achieved in the 2kx2k image in Figure 5. A detailed discussion of this classification subsystem and its utility for road network extraction can be found in Riseman et at. 1997) . To further improve performance in terrain classification, we have employed information from 3D reconstruction of the Terrest module. In particular, a new concept of 3D texture has been recently proposed (Wang et al. 1997a, Wang et al. 1997b) increasing the dimensionality in texture analysis. ....

....of the search range, then the SF is considered well defined. Altogether four 3D features have been developed from the statistical properties of these similarity functions. The characteristics of the features are summarized in Table 1. Details of the generation of these 3D features can be found in Wang et al. 1997a,b) Experiments have tested the utility of 3D texture features in terrain classification, by themselves and in conjunction with traditional 2D feature sets. Figure 3 shows a stereo pair used in one such test evaluation. The images were 2k Theta 2k sections extracted from aerial images of a ....

[Article contains additional citation context not shown here]

In: R. Howard and J. Matheson (eds), Readings on the Principles and Applications of Decision Analysis, Vol. 2, Menlo Park, pp. 721--762. Jaynes, C. O., Marengoni, M., Hanson, A. and Riseman, E., 1997. Automatic model acquisition through intelligent control.

No context found.

Howard, R. A.; Matheson, J. E., eds. Readings on the Principles and Applications of Decision Analysis. Menlo Park, CA: Strategic Decisions Group; 1983.

No context found.

In Howard, R. and Matheson, J., editors, Readings on the Principles and Applications of Decision Analysis, volume II, pages 721--762. Strategic Decisions Group, Menlo Park, CA. Laskey, K. B. (1990). Adapting connectionist learning to Bayes networks. International Journal of Approximate Reasoning, 4:261--282.

First 50 documents

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