Ulrich Junker - Academia.edu (original) (raw)

Papers by Ulrich Junker

Dagstuhl Seminar Proceedings, 2006

Combinatorial problems such as scheduling, resource allocation, and configuration have many attri... more Combinatorial problems such as scheduling, resource allocation, and configuration have many attributes that can be subject of user preferences. Traditional optimization approaches compile those preferences into a single utility function and use it as the optimization objective when solving the problem, but neither explain why the resulting solution satisfies the original preferences, nor indicate the trade-offs made during problem solving. We argue that the whole problem solving process becomes more transparent and controllable for the user if it is based on the original preferences. We show how the original preferences can be used to control the problem solving process and how they can be used to explain the choice and the optimality of the detected solution. Based on this explanation, the user can refine the preference model, thus gaining full control over the problem solver.

National Conference on Artificial Intelligence, Jul 28, 2002

Many real-world AI problems (e.g. in configuration) are weakly constrained, thus requiring a mech... more Many real-world AI problems (e.g. in configuration) are weakly constrained, thus requiring a mechanism for characterizing and finding the preferred solutions. Preferencebased search (PBS) exploits preferences between decisions to focus search to preferred solutions, but does not efficiently treat preferences on defined criteria such as the total price or quality of a configuration. We generalize PBS to compute balanced, extreme, and Pareto-optimal solutions for general CSP's, thus handling preferences on and between multiple criteria. A master-PBS selects criteria based on trade-offs and preferences and passes them as optimization objective to a sub-PBS that performs a constraint-based Branch-and-Bound search. We project the preferences of the selected criterion to the search decisions to provide a search heuristics and to reduce search effort, thus giving the criterion a high impact on the search. The resulting method will particularly be effective for CSP's with large domains that arise if configuration catalogs are large.

National Conference on Artificial Intelligence, Jul 30, 2000

Preference-based search (PBS) is a new search procedure for solving combinatorial optimization pr... more Preference-based search (PBS) is a new search procedure for solving combinatorial optimization problems. Given a set of preferences between search decisions, PBS searches through a space of preferred solutions, which is tighter than the space of all solutions. The definition of preferred solutions is based on work in non-monotonic reasoning (Brewka 1989; Geffner & Pearl 1992; Grosof 1991) on priorities between defaults. The basic idea of PBS is quite simple: Always pick a locally best decision α. Either make the decision α or make other locally best decisions that allow to deduce ¬α and thus represent a counterargument for α. If there is no possible counterargument then PBS does not explore the subtree of ¬α. This pruning of the search space is obtained by nonmonotonic inference rules that are inspired by Doyle's TMS and that detect decisions belonging to all or no preferred solution. We show that PBS can optimally solve various important scheduling problems.

Artificial intelligence for engineering design, analysis and manufacturing, Feb 1, 2003

Preference programming provides a new paradigm for expressing (default) decisions, preferences be... more Preference programming provides a new paradigm for expressing (default) decisions, preferences between decisions, and search strategies in a declarative and unified way and for embedding them in a constraint and rule language. Business experts can thus directly specify preferences and search directives in form of rules without needing to program search strategies as required by constraint programming based configuration tools. Preference programming allows to describe preferences between individual decisions, as well as groups of decisions and decision rules. There can be dynamic (or context-dependent) preferences, inconsistent preferences, and meta-preferences. Following [Brewka, 1989; Junker, 1997], preferences constrain the order in which decisions are made during search. It is possible to enumerate all configurations or to focus search to preferred configurations, which respect the default choices and preferences of the user.

Springer eBooks, Jan 5, 2009

Partial orders provide a convenient way to express preferences on multiple criteria. Prominent ex... more Partial orders provide a convenient way to express preferences on multiple criteria. Prominent examples are Pareto-dominance and the preference relations of (T)CP-nets [1]. In advanced personalized recommender systems, the user may also specify a partial order over the possible values of a single criterion. We introduce a technique called outer branching to compute the non-dominated frontier of optimization problems with partial orders. It can be used to compute all Pareto-optimal solutions for n criteria by performing a systematic search over the criteria space. Dominance constraints avoid the generation of non-optimal solutions.

International Joint Conference on Artificial Intelligence, Aug 9, 2003

We present an algorithm Pref-AC that limits arc consistency (AC) to the preferred choices of a tr... more We present an algorithm Pref-AC that limits arc consistency (AC) to the preferred choices of a tree search procedure and that makes constraint solving more efficient without changing the pruning and shape of the search tree. Arc consistency thus becomes more scalable and usable for many realworld constraint satisfaction problems such as configuration and scheduling. Moreover, Pref-AC directly computes a preferred solution for tree-like constraint satisfaction problems.

Annals of Mathematics and Artificial Intelligence, Mar 1, 1994

Preferences between diagnostic assumptions are needed to handle interactions between different ki... more Preferences between diagnostic assumptions are needed to handle interactions between different kinds of assumptions and to focus the diagnostic process to components that are more likely to fail. We investigate different preference criteria and relate them to search strategies in Reiter's hitting trees. In particular, we consider a partial order on assumptions. 9 Difference in robustness: If some components are more likely to fail due to their physical properties then these components should be inspected first. Therefore, it is sufficient to generate diagnoses containing likely faults in the beginning. For example, if the correctness assumptions of a wire, a bulb, and a resistor are in conflict then we only retract that of the less robust bulb. 9 Interaction of assumptions: In some approaches [1, 13], correctness and fault assumptions are used in combination, which can lead to unintended interactions. Normally, one wants to prefer the correctness assumptions. For example, if we do not observe light in our room we first think that the switch is off instead of assuming that the bulb or something else is broken. 9 Description levels: Technical systems are usually described on different abstraction levels, which enables a top-down-refinement strategy for finding faults. However, the introduction of assumptions has to be controlled in order to realize this strategy. An assumption of a component of a subsystem should only be introduced if there is a reason that this subsystem is defective.

Springer eBooks, 2008

Combinatorial problems such as scheduling, resource allocation, and configuration have many attri... more Combinatorial problems such as scheduling, resource allocation, and configuration have many attributes that can be subject of user preferences. Traditional optimization approaches compile those preferences into a single utility function and use it as the optimization objective when solving the problem, but neither explain why the resulting solution satisfies the original preferences, nor indicate the trade-offs made during problem solving. We argue that the whole problem solving process becomes more transparent and controllable for the user if it is based on the original preferences. We show how the original preferences can be used to control the problem solving process and how they can be used to explain the choice and the optimality of the detected solution. Based on this explanation, the user can refine the preference model, thus gaining full control over the problem solver.

International Joint Conference on Artificial Intelligence, Aug 20, 1995

Dynamic objects such as liquids, waves, and flames can easily change their position, shape, and n... more Dynamic objects such as liquids, waves, and flames can easily change their position, shape, and number. Snapshot images produced by finite element simulators show these changes, hut lack an explicit representation of the objects and their causes. For the example of seismic waves, we develop a method for interpreting snapshots which is based on Hayes 7 concept of a history.

National Conference on Artificial Intelligence, Jul 29, 1990

In this pa.per we develop a proof procedure for autoepistemic (AEL) and defalrlt logics (DL), bas... more In this pa.per we develop a proof procedure for autoepistemic (AEL) and defalrlt logics (DL), based on translating them into a. Truth Maintenance System (TMS). The translation is decidable if t,he theory consists of a finite number of defaults a.nd premises and classical derivability for the base language is decida.ble. To determine all extensions of a network, we develop variants of Doyle's labelling algorithms.

Lecture Notes in Computer Science, 1991

Without Abstract

National Conference on Artificial Intelligence, Jul 25, 2004

Over-constrained problems can have an exponential number of conflicts, which explain the failure,... more Over-constrained problems can have an exponential number of conflicts, which explain the failure, and an exponential number of relaxations, which restore the consistency. A user of an interactive application, however, desires explanations and relaxations containing the most important constraints. To address this need, we define preferred explanations and relaxations based on user preferences between constraints and we compute them by a generic method which works for arbitrary CP, SAT, or DL solvers. We significantly accelerate the basic method by a divide-and-conquer strategy and thus provide the technological basis for the explanation facility of a principal industrial constraint programming tool, which is, for example, used in numerous configuration applications.

Lecture Notes in Computer Science, 1998

Finding good problem decompositions is crucial for solving large-scale key/lock configuration pro... more Finding good problem decompositions is crucial for solving large-scale key/lock configuration problems. We present a novel approach to problem decomposition where the detection of a subproblem hierarchy is formulated as a constraint satisfaction problem (CSP) on set variables. Primitive constraints on set variables such as an inverse and a union-over-set-constraint are used to formalize properties of trees and to ensure that solutions of these CSPs represent a tree. An objective for optimizing properties of the tree is presented as well. Experimental results from an industrial prototype are reported

International Joint Conference on Artificial Intelligence, Aug 23, 1997

Explicit preferences on assumptions as used in prioritized circumscription [McCarthy, 1986; Lifsc... more Explicit preferences on assumptions as used in prioritized circumscription [McCarthy, 1986; Lifschitz, 1985; Grosof, 1991] and preferred subtheories [Brewka, 1989] provide a clear and declarative method for defining preferred models. In this paper, we show how to embed preferences in the logical theory itself. This gives a high freedom for expressing statements about preferences. Preferences can now depend on other assumptions and are thus dynamic. We elaborate a preferential semantics based on Lehmann's cumulative models, as well as a corresponding constructive characterization, which specifies how to correctly treat dynamic preferences in the default reasoning system EXCEPT [Junker, 1992].

Lecture Notes in Computer Science, 2009

We consider constructive approaches to decision making which allow incomplete preference orders o... more We consider constructive approaches to decision making which allow incomplete preference orders over multiple criteria. Whereas additional preferences may be acquired during the decision making process, the set of criteria is usually kept fixed. In this paper, we study the addition of new criteria and examine how this may refine or even reverse the existing preferences. We identify essential changes in the preference order and show that these changes provide a compact representation of preference relations in an open world.

International Joint Conference on Artificial Intelligence, Aug 24, 1991

We demonstrate the technological value of nonmonotonic logics by an example: We use prioritized d... more We demonstrate the technological value of nonmonotonic logics by an example: We use prioritized defaults for candidate generation in diagnosis from first principles. We implement this non-monotonic logic by TMS similar to default logic. Prioritized defaults allow an easy formu lation of a diagnosis problem including state ments such as 'eletrical parts are more reliable than mechanical ones' or 'prefer correct mod els to fault models' since defaults are put into different levels of reliability. These preferences prune some counterarguments in TMS and thus lead to a reduced network. Moreover, the labelings of this network are exactly the preferred subtheories of its prioritized default theory.

European Conference on Artificial Intelligence, May 22, 2006

Inconsistency proving of CSPs is typically achieved by a combination of systematic search and arc... more Inconsistency proving of CSPs is typically achieved by a combination of systematic search and arc consistency, which can both be characterized as resolution. However, it is well-known that there are cases where resolution produces exponential contradiction proofs, although proofs of polynomial size exist. For this reason, we will use optimization methods to reduce the proof size globally by 1. decomposing

Knowledge Engineering Review, Jul 26, 2012

The European air traffic flow management problem poses particular challenges on optimization tech... more The European air traffic flow management problem poses particular challenges on optimization technology as it requires detailed modelling and rapid online optimization capabilities. Constraint programming proved successful in addressing these challenges for departure time slot allocation by offering fine-grained modelling of resource constraints and fast allocation through heuristic-repair strategies.

Foundations of artificial intelligence, 2006

Annals of Operations Research, Aug 1, 2004

Many real-world AI problems (e.g. in configuration) are weakly constrained, thus requiring a mech... more Many real-world AI problems (e.g. in configuration) are weakly constrained, thus requiring a mechanism for characterizing and finding the preferred solutions. Preferencebased search (PBS) exploits preferences between decisions to focus search to preferred solutions, but does not efficiently treat preferences on defined criteria such as the total price or quality of a configuration. We generalize PBS to compute balanced, extreme, and Pareto-optimal solutions for general CSP's, thus handling preferences on and between multiple criteria. A master-PBS selects criteria based on trade-offs and preferences and passes them as optimization objective to a sub-PBS that performs a constraint-based Branch-and-Bound search. We project the preferences of the selected criterion to the search decisions to provide a search heuristics and to reduce search effort, thus giving the criterion a high impact on the search. The resulting method will particularly be effective for CSP's with large domains that arise if configuration catalogs are large.

Dagstuhl Seminar Proceedings, 2006

Combinatorial problems such as scheduling, resource allocation, and configuration have many attri... more Combinatorial problems such as scheduling, resource allocation, and configuration have many attributes that can be subject of user preferences. Traditional optimization approaches compile those preferences into a single utility function and use it as the optimization objective when solving the problem, but neither explain why the resulting solution satisfies the original preferences, nor indicate the trade-offs made during problem solving. We argue that the whole problem solving process becomes more transparent and controllable for the user if it is based on the original preferences. We show how the original preferences can be used to control the problem solving process and how they can be used to explain the choice and the optimality of the detected solution. Based on this explanation, the user can refine the preference model, thus gaining full control over the problem solver.

National Conference on Artificial Intelligence, Jul 28, 2002

Many real-world AI problems (e.g. in configuration) are weakly constrained, thus requiring a mech... more Many real-world AI problems (e.g. in configuration) are weakly constrained, thus requiring a mechanism for characterizing and finding the preferred solutions. Preferencebased search (PBS) exploits preferences between decisions to focus search to preferred solutions, but does not efficiently treat preferences on defined criteria such as the total price or quality of a configuration. We generalize PBS to compute balanced, extreme, and Pareto-optimal solutions for general CSP's, thus handling preferences on and between multiple criteria. A master-PBS selects criteria based on trade-offs and preferences and passes them as optimization objective to a sub-PBS that performs a constraint-based Branch-and-Bound search. We project the preferences of the selected criterion to the search decisions to provide a search heuristics and to reduce search effort, thus giving the criterion a high impact on the search. The resulting method will particularly be effective for CSP's with large domains that arise if configuration catalogs are large.

National Conference on Artificial Intelligence, Jul 30, 2000

Preference-based search (PBS) is a new search procedure for solving combinatorial optimization pr... more Preference-based search (PBS) is a new search procedure for solving combinatorial optimization problems. Given a set of preferences between search decisions, PBS searches through a space of preferred solutions, which is tighter than the space of all solutions. The definition of preferred solutions is based on work in non-monotonic reasoning (Brewka 1989; Geffner & Pearl 1992; Grosof 1991) on priorities between defaults. The basic idea of PBS is quite simple: Always pick a locally best decision α. Either make the decision α or make other locally best decisions that allow to deduce ¬α and thus represent a counterargument for α. If there is no possible counterargument then PBS does not explore the subtree of ¬α. This pruning of the search space is obtained by nonmonotonic inference rules that are inspired by Doyle's TMS and that detect decisions belonging to all or no preferred solution. We show that PBS can optimally solve various important scheduling problems.

Artificial intelligence for engineering design, analysis and manufacturing, Feb 1, 2003

Preference programming provides a new paradigm for expressing (default) decisions, preferences be... more Preference programming provides a new paradigm for expressing (default) decisions, preferences between decisions, and search strategies in a declarative and unified way and for embedding them in a constraint and rule language. Business experts can thus directly specify preferences and search directives in form of rules without needing to program search strategies as required by constraint programming based configuration tools. Preference programming allows to describe preferences between individual decisions, as well as groups of decisions and decision rules. There can be dynamic (or context-dependent) preferences, inconsistent preferences, and meta-preferences. Following [Brewka, 1989; Junker, 1997], preferences constrain the order in which decisions are made during search. It is possible to enumerate all configurations or to focus search to preferred configurations, which respect the default choices and preferences of the user.

Springer eBooks, Jan 5, 2009

Partial orders provide a convenient way to express preferences on multiple criteria. Prominent ex... more Partial orders provide a convenient way to express preferences on multiple criteria. Prominent examples are Pareto-dominance and the preference relations of (T)CP-nets [1]. In advanced personalized recommender systems, the user may also specify a partial order over the possible values of a single criterion. We introduce a technique called outer branching to compute the non-dominated frontier of optimization problems with partial orders. It can be used to compute all Pareto-optimal solutions for n criteria by performing a systematic search over the criteria space. Dominance constraints avoid the generation of non-optimal solutions.

International Joint Conference on Artificial Intelligence, Aug 9, 2003

We present an algorithm Pref-AC that limits arc consistency (AC) to the preferred choices of a tr... more We present an algorithm Pref-AC that limits arc consistency (AC) to the preferred choices of a tree search procedure and that makes constraint solving more efficient without changing the pruning and shape of the search tree. Arc consistency thus becomes more scalable and usable for many realworld constraint satisfaction problems such as configuration and scheduling. Moreover, Pref-AC directly computes a preferred solution for tree-like constraint satisfaction problems.

Annals of Mathematics and Artificial Intelligence, Mar 1, 1994

Preferences between diagnostic assumptions are needed to handle interactions between different ki... more Preferences between diagnostic assumptions are needed to handle interactions between different kinds of assumptions and to focus the diagnostic process to components that are more likely to fail. We investigate different preference criteria and relate them to search strategies in Reiter's hitting trees. In particular, we consider a partial order on assumptions. 9 Difference in robustness: If some components are more likely to fail due to their physical properties then these components should be inspected first. Therefore, it is sufficient to generate diagnoses containing likely faults in the beginning. For example, if the correctness assumptions of a wire, a bulb, and a resistor are in conflict then we only retract that of the less robust bulb. 9 Interaction of assumptions: In some approaches [1, 13], correctness and fault assumptions are used in combination, which can lead to unintended interactions. Normally, one wants to prefer the correctness assumptions. For example, if we do not observe light in our room we first think that the switch is off instead of assuming that the bulb or something else is broken. 9 Description levels: Technical systems are usually described on different abstraction levels, which enables a top-down-refinement strategy for finding faults. However, the introduction of assumptions has to be controlled in order to realize this strategy. An assumption of a component of a subsystem should only be introduced if there is a reason that this subsystem is defective.

Springer eBooks, 2008

Combinatorial problems such as scheduling, resource allocation, and configuration have many attri... more Combinatorial problems such as scheduling, resource allocation, and configuration have many attributes that can be subject of user preferences. Traditional optimization approaches compile those preferences into a single utility function and use it as the optimization objective when solving the problem, but neither explain why the resulting solution satisfies the original preferences, nor indicate the trade-offs made during problem solving. We argue that the whole problem solving process becomes more transparent and controllable for the user if it is based on the original preferences. We show how the original preferences can be used to control the problem solving process and how they can be used to explain the choice and the optimality of the detected solution. Based on this explanation, the user can refine the preference model, thus gaining full control over the problem solver.

International Joint Conference on Artificial Intelligence, Aug 20, 1995

Dynamic objects such as liquids, waves, and flames can easily change their position, shape, and n... more Dynamic objects such as liquids, waves, and flames can easily change their position, shape, and number. Snapshot images produced by finite element simulators show these changes, hut lack an explicit representation of the objects and their causes. For the example of seismic waves, we develop a method for interpreting snapshots which is based on Hayes 7 concept of a history.

National Conference on Artificial Intelligence, Jul 29, 1990

In this pa.per we develop a proof procedure for autoepistemic (AEL) and defalrlt logics (DL), bas... more In this pa.per we develop a proof procedure for autoepistemic (AEL) and defalrlt logics (DL), based on translating them into a. Truth Maintenance System (TMS). The translation is decidable if t,he theory consists of a finite number of defaults a.nd premises and classical derivability for the base language is decida.ble. To determine all extensions of a network, we develop variants of Doyle's labelling algorithms.

Lecture Notes in Computer Science, 1991

Without Abstract

National Conference on Artificial Intelligence, Jul 25, 2004

Over-constrained problems can have an exponential number of conflicts, which explain the failure,... more Over-constrained problems can have an exponential number of conflicts, which explain the failure, and an exponential number of relaxations, which restore the consistency. A user of an interactive application, however, desires explanations and relaxations containing the most important constraints. To address this need, we define preferred explanations and relaxations based on user preferences between constraints and we compute them by a generic method which works for arbitrary CP, SAT, or DL solvers. We significantly accelerate the basic method by a divide-and-conquer strategy and thus provide the technological basis for the explanation facility of a principal industrial constraint programming tool, which is, for example, used in numerous configuration applications.

Lecture Notes in Computer Science, 1998

Finding good problem decompositions is crucial for solving large-scale key/lock configuration pro... more Finding good problem decompositions is crucial for solving large-scale key/lock configuration problems. We present a novel approach to problem decomposition where the detection of a subproblem hierarchy is formulated as a constraint satisfaction problem (CSP) on set variables. Primitive constraints on set variables such as an inverse and a union-over-set-constraint are used to formalize properties of trees and to ensure that solutions of these CSPs represent a tree. An objective for optimizing properties of the tree is presented as well. Experimental results from an industrial prototype are reported

International Joint Conference on Artificial Intelligence, Aug 23, 1997

Explicit preferences on assumptions as used in prioritized circumscription [McCarthy, 1986; Lifsc... more Explicit preferences on assumptions as used in prioritized circumscription [McCarthy, 1986; Lifschitz, 1985; Grosof, 1991] and preferred subtheories [Brewka, 1989] provide a clear and declarative method for defining preferred models. In this paper, we show how to embed preferences in the logical theory itself. This gives a high freedom for expressing statements about preferences. Preferences can now depend on other assumptions and are thus dynamic. We elaborate a preferential semantics based on Lehmann's cumulative models, as well as a corresponding constructive characterization, which specifies how to correctly treat dynamic preferences in the default reasoning system EXCEPT [Junker, 1992].

Lecture Notes in Computer Science, 2009

We consider constructive approaches to decision making which allow incomplete preference orders o... more We consider constructive approaches to decision making which allow incomplete preference orders over multiple criteria. Whereas additional preferences may be acquired during the decision making process, the set of criteria is usually kept fixed. In this paper, we study the addition of new criteria and examine how this may refine or even reverse the existing preferences. We identify essential changes in the preference order and show that these changes provide a compact representation of preference relations in an open world.

International Joint Conference on Artificial Intelligence, Aug 24, 1991

We demonstrate the technological value of nonmonotonic logics by an example: We use prioritized d... more We demonstrate the technological value of nonmonotonic logics by an example: We use prioritized defaults for candidate generation in diagnosis from first principles. We implement this non-monotonic logic by TMS similar to default logic. Prioritized defaults allow an easy formu lation of a diagnosis problem including state ments such as 'eletrical parts are more reliable than mechanical ones' or 'prefer correct mod els to fault models' since defaults are put into different levels of reliability. These preferences prune some counterarguments in TMS and thus lead to a reduced network. Moreover, the labelings of this network are exactly the preferred subtheories of its prioritized default theory.

European Conference on Artificial Intelligence, May 22, 2006

Inconsistency proving of CSPs is typically achieved by a combination of systematic search and arc... more Inconsistency proving of CSPs is typically achieved by a combination of systematic search and arc consistency, which can both be characterized as resolution. However, it is well-known that there are cases where resolution produces exponential contradiction proofs, although proofs of polynomial size exist. For this reason, we will use optimization methods to reduce the proof size globally by 1. decomposing

Knowledge Engineering Review, Jul 26, 2012

The European air traffic flow management problem poses particular challenges on optimization tech... more The European air traffic flow management problem poses particular challenges on optimization technology as it requires detailed modelling and rapid online optimization capabilities. Constraint programming proved successful in addressing these challenges for departure time slot allocation by offering fine-grained modelling of resource constraints and fast allocation through heuristic-repair strategies.

Foundations of artificial intelligence, 2006

Annals of Operations Research, Aug 1, 2004

Many real-world AI problems (e.g. in configuration) are weakly constrained, thus requiring a mech... more Many real-world AI problems (e.g. in configuration) are weakly constrained, thus requiring a mechanism for characterizing and finding the preferred solutions. Preferencebased search (PBS) exploits preferences between decisions to focus search to preferred solutions, but does not efficiently treat preferences on defined criteria such as the total price or quality of a configuration. We generalize PBS to compute balanced, extreme, and Pareto-optimal solutions for general CSP's, thus handling preferences on and between multiple criteria. A master-PBS selects criteria based on trade-offs and preferences and passes them as optimization objective to a sub-PBS that performs a constraint-based Branch-and-Bound search. We project the preferences of the selected criterion to the search decisions to provide a search heuristics and to reduce search effort, thus giving the criterion a high impact on the search. The resulting method will particularly be effective for CSP's with large domains that arise if configuration catalogs are large.