Michael C. Horsch

Agent Reasoning Mechanism for Long–Term Coalitions Based on Decision Making and Trust

We address long–term coalitions that are formed of both customer and vendor agents. We present a coalition formation mechanism designed at the agent level as a decision problem. The proposed mechanism is analyzed at both system and agent levels. Our results show that the coalition formation mechanism is beneficial for both the system—it reaches an equilibrium state—and for the agents—their gains highly increase over time.

Flexible policy construction by information refinement

We report on work towards flexible algorithms for solving decision problems represented as influence diagrams. An algorithm is given to construct a tree structure for each decision node in an influence diagram. Each tree represents a decision function and is constructed incrementally. The improvements to the tree converge to the optimal decision function (neglecting computational costs) and the asymptotic behaviour is only a constant factor worse than dynamic programming techniques, counting the number of Bayesian network queries. Empirical results show how expected utility increases with the size of the tree and the number of Bayesian net calculations.

Conditioning graphs: practical structures for inference in bayesian networks

Programmers employing inference in Bayesian networks typically rely on the inclusion of the model as well as an inference engine into their application. Sophisticated inference engines require non-trivial amounts of space and are also difficult to implement. This limits their use in some applications that would otherwise benefit from probabilistic inference. This paper presents a system that minimizes the space requirement of the model. The inference engine is sufficiently simple as to avoid space-limitation and be easily implemented in almost any environment. We show a fast, compact indexing structure that is linear in the size of the network. The additional space required to compute over the model is linear in the number of variables in the network.

An hierarchical terrain representation for approximately shortest paths

We propose a fast algorithm for on-line path search in grid-like undirected planar graphs with real edge costs (aka terrains). Our algorithm depends on an off-line analysis of the graph, requiring poly-logarithmic time and space. The off-line preprocessing constructs a hierarchical representation which allows detection of features specific to the terrain. While our algorithm is not guaranteed to find an optimal path, we demonstrate empirically that it is very fast, and that the difference from optimal is almost always small.

Methods for constructing balanced elimination trees and other recursive decompositions

An elimination tree is a form of recursive factorization for Bayesian networks. Elimination trees can be used as the basis for a practical implementation of Bayesian network inference via conditioning graphs. The time complexity for inference in elimination trees has been shown to be O(nexp(d)), where d is the height of the elimination tree. In this paper, we demonstrate two new heuristics for building small elimination trees. We also demonstrate a simple technique for deriving elimination trees from Darwiche et al.s dtrees, and vice versa. We show empirically that our heuristics, combined with a constructive process for building elimination trees, produces the smaller elimination trees than previous methods.

Exploiting dynamic independence in a static conditioning graph

A conditioning graph (CG) is a graphical structure that attempt to minimize the implementation overhead of computing probabilities in belief networks. A conditioning graph recursively factorizes the network, but restricting each decomposition to a single node allows us to store the structure with minimal overhead, and compute with a simple algorithm. This paper extends conditioning graphs with optimizations that effectively reduce the height of the CG, thus reducing time complexity exponentially, while increasing the storage requirements by only a constant factor. We conclude that CGs are frequently as efficient as any other exact inference method, with the advantage of being vastly superior to VE and JT in terms of space complexity, and far simpler to implement.

A decision theoretic meta-reasoner for constraint optimization

Solving constraint optimization problems is hard because it is not enough to find the best solution; an algorithm does not know a candidate is the best solution until it has proven that there are no better solutions The proof can be long, compared to the time spent to find a good solution In the cases where there are resource bounds, the proof of optimality may not be achievable and a tradeoff needs to be made between the solution quality and the cost due to the time delay We propose a decision theoretic meta-reasoning-guided COP solver to address this issue By choosing the action with the estimated maximal expected utility, the meta-reasoner finds a stopping point with a good tradeoff between the solution quality and the time cost.

Generalized Arc Consistency with Application to MaxCSP

We present an abstract generalization of arc consistency which subsumes the definition of arc consistency in classical CSPs. Our generalization is based on the view of local consistency as technique for approximation of marginal solutions. These approximations are intended for use as heuristics during search.We show that this generalization leads to useful application in classical CSPs as well as non-classical CSPs such as MaxCSP, and instances of the Semi-ring CSP formalism developed by Bistarelli et al. [2]. We demonstrate the application ofthe theory by developing a novel algorithm for use in solving MaxCSP.

Predicting good propagation methods for constraint satisfaction

Given the breadth of constraint satisfaction problems (CSPs) and the wide variety of CSP solvers, it can be difficult to determine a priori which solving method is best suited to a problem. We explore the use of machine learning to predict which solving method will be most effective for a given problem. Our investigation studies the problem of attribute selection for CSPs, and supervised learning to classify CSP instances drawn from four distinct CSP classes. We limit our study to the choice of two well-known, but simple, CSP solvers. We show that the average performance of the resulting solver is very close to the average performance of a CSP solver based on an oracle.

Camera selection using SCSPs

An automated director is needed for sports video games to select between multiple camera views. An SCSP approach enables setting preferences for views that depend on the current situation. This approach is better than using a classical CSP, or a finite state machine.