Christian Eichhorn

Measuring flow as concept for detecting game fun in the Pac-Man game

Popular games often have a high-quality graphic design but quite simple-minded non player characters (NPC). Recently, Computational Intelligence (CI) methods have been discovered as suitable methods to revive NPC, making games more interesting, challenging, and funny. We present a fairly large study of human players on the simple arcade game Pac-Man, controlling the ghosts behaviors by simple strategies, neural networks or evolutionary algorithms. The playerpsilas fun is of course a subjective experience, but we presume that it is related to the psychological flow concept. We deal with the question whether flow is a more reliable measure than asking human players directly for the fun experienced during the game. In order to detect flow, we introduce a measure based on the interaction time fraction between the human-controlled Pac-Man and the ghosts, and compare the outcome to the results of a fun measure suggested by Yannakakis and Hallam [1].

Structural Inference from Conditional Knowledge Bases

There are several approaches implementing reasoning based on conditional knowledge bases, one of the most popular being System Z (Pearl, Proceedings of the 3rd conference on theoretical aspects of reasoning about knowledge, TARK ’90, Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, pp. 121–135, 1990). We look at ranking functions (Spohn, The Laws of Belief: Ranking Theory and Its Philosophical Applications, Oxford University Press, Oxford, 2012) in general, conditional structures and c-representations (Kern-Isberner, Conditionals in Nonmonotonic Reasoning and Belief Revision: Considering Conditionals as Agents, vol. 2087 of LNCS, Springer, Berlin, 2001) in order to examine the reasoning strength of the different approaches by learning which of the known calculi of nonmonotonic reasoning (System P and R) and Direct Inference are applicable to these inference relations. Furthermore we use the recently proposed Enforcement-postulate (Kern-Isberner and Krümpelmann, Proceedings of the 22nd international joint conference on artificial intelligence, vol. 2, IJCAI’11, AAAI Press, pp. 937–942, 2011) to show dependencies between these approaches.

Skeptical Inference Based on C-Representations and Its Characterization as a Constraint Satisfaction Problem

The axiomatic system P is an important standard for plausible, nonmonotonic inferences that is, however, known to be too weak to solve benchmark problems like irrelevance, or subclass inheritance so-called Drowning Problem. Spohns ranking functions which provide a semantic base for system P have often been used to design stronger inference relations, like Pearls system Z, or c-representations. While each c-representation shows excellent inference properties and handles particularly irrelevance and subclass inheritance properly, it is still an open problem which c-representation is the best. In this paper, we focus on the generic properties of c-representations and consider the skeptical inference relation c-inference that is obtained by taking all c-representations of a given knowledge base into account. In particular, we show that c-inference preserves the properties of solving irrelevance and subclass inheritance which are met by every single c-representation. Moreover, we characterize skeptical c-inference as a constraint satisfaction problem so that constraint solvers can be used for its implementation.

Semantic duplicate identification with parsing and machine learning

Identifying duplicate texts is important in many areas like plagiarism detection, information retrieval, text summarization, and question answering. Current approaches are mostly surface-oriented (or use only shallow syntactic representations) and see each text only as a token list. In this work however, we describe a deep, semantically oriented method based on semantic networks which are derived by a syntactico-semantic parser. Semantically identical or similar semantic networks for each sentence of a given base text are efficiently retrieved by using a specialized index. In order to detect many kinds of paraphrases the semantic networks of a candidate text are varied by applying inferences: lexico-semantic relations, relation axioms, and meaning postulates. Important phenomena occurring in difficult duplicates are discussed. The deep approach profits from background knowledge, whose acquisition from corpora is explained briefly. The deep duplicate recognizer is combined with two shallow duplicate recognizers in order to guarantee a high recall for texts which are not fully parsable. The evaluation shows that the combined approach preserves recall and increases precision considerably in comparison to traditional shallow methods.

Using inductive reasoning for completing OCF-networks

OCF-networks provide the possibility to combine qualitative information expressed by rankings of (conditional) formulas with the strong structural information of a network, in this respect being a qualitative variant of the better known Bayesian networks. Like for Bayesian networks, a global ranking function can be calculated quickly and efficiently from the locally distributed information, whereas the latter significantly reduces the exponentially high complexity of the semantical ranking approach. This qualifies OCF-networks for applications. However, in practical applications the provided ranking information may not be in the format needed to be represented by an OCF-network, or some values may be simply missing. In this paper, we present techniques for filling in the missing values using methods of inductive reasoning and we elaborate on formal properties of OCF-networks.

LEG Networks for Ranking Functions

When using representations of plausibility for semantical frameworks, the storing capacity needed is usually exponentially in the number of variables. Therefore, network-based approaches that decompose the semantical space have proven to be fruitful in environments with probabilistic information. For applications where a more qualitative information is preferable to quantitative information, ordinal conditional functions (OCF) offer a convenient methodology. Here, Bayesian-like networks have been proposed for ranking functions, so called OCF-networks. These networks not only suffer from similar problems as Bayesian networks, in particular, allowing only restricted classes of conditional relationships, it also has been found recently that problems with admissibility may arise. In this paper we propose LEG networks for ranking functions, also carrying over an idea from probabilistics. OCF-LEG networks can be built for any conditional knowledge base and filled by local OCF that can be found by inductive reasoning. A global OCF is set up from the local ones, and it is shown that the global OCF is admissible with respect to the underlying knowledge base.

A Practical Comparison of Qualitative Inferences with Preferred Ranking Models

When reasoning qualitatively from a conditional knowledge base, two established approaches are system Z and p-entailment. The latter infers skeptically over all ranking models of the knowledge base, while system Z uses the unique pareto-minimal ranking model for the inference relations. Between these two extremes of using all or just one ranking model, the approach of c-representations generates a subset of all ranking models with certain constraints. Recent work shows that skeptical inference over all c-representations of a knowledge base includes and extends p-entailment. In this paper, we follow the idea of using preferred models of the knowledge base instead of the set of all models as a base for the inference relation. We employ different minimality constraints for c-representations and demonstrate inference relations from sets of preferred c-representations with respect to these constraints. We present a practical tool for automatic c-inference that is based on a high-level, declarative constraint-logic programming approach. Using our implementation, we illustrate that different minimality constraints lead to inference relations that differ mutually as well as from system Z and p-entailment.

CP- and OCF-networks - a comparison

Network approaches are used to structure, partition and display formalisms in the area of knowledge representation as well as decision making. Known approaches are, for instance, OCF-networks, Bayesian style networks where every variable is annotated with a conditional ranking table, and CP-networks, directed acyclic networks with local preferences annotated at each vertex. The structures of these networks are similar, but their semantics seem to be quite different. In this paper we discuss if OCF-networks can be used to model the information of CP-networks and vice versa. To answer this question we investigate which restrictions and conditions have to be presupposed to either of the approaches such that one structure can be used to generate the other.

Properties of skeptical c-inference for conditional knowledge bases and its realization as a constraint satisfaction problem

While the axiomatic system P is an important standard for plausible, nonmonotonic inferences from conditional knowledge bases, it is known to be too weak to solve benchmark problems like Irrelevance or Subclass Inheritance. Ordinal conditional functions provide a semantic base for system P and have often been used to design stronger inference relations, like Pearl’s system Z, or c-representations. While each c-representation shows excellent inference properties and handles particularly Irrelevance and Subclass Inheritance properly, it is still an open problem which c-representation is the best. In this paper, we consider the skeptical inference relation, called c-inference, that is obtained by taking all c-representations of a given knowledge base into account. We study properties of c-inference and show in particular that it preserves the properties of solving Irrelevance and Subclass Inheritance. Based on a characterization of c-representations as solutions of a Constraint Satisfaction Problem (CSP), we also model skeptical c-inference as a CSP and prove soundness and completeness of the modelling, ensuring that constraint solvers can be used for implementing c-inference.

On Transformations and Normal Forms of Conditional Knowledge Bases.

Background knowledge is often represented by sets of conditionals of the form “if A then usually B”. Such knowledge bases should not be circuitous, but compact and easy to compare in order to allow for efficient processing in approaches dealing with and inferring from background knowledge, such as nonmonotonic reasoning. In this paper we present transformation systems on conditional knowledge bases that allow to identify and remove unnecessary conditionals from the knowledge base while preserving the knowledge base’s model set.