Jörg Denzinger

Evaluating Different Genetic Operators in the Testing for Unwanted Emergent Behavior Using Evolutionary Learning of Behavior

We present an experimental comparison of different genetic operators regarding their use in an evolutionary learning method that searches for unwanted emergent behavior in a multi-agent system. The idea of the learning method is to evolve cooperative behavior of a group of so-called attack agents that act in the same environment as the tested agents. The attack agents use action sequences as agent architecture and the quality of a group of such agents is measured by how near their behavior brings the tested agents to show the unwanted behavior. Our experiments within the ARES II rescue simulator with an agent team written by students show that this method is able to find unwanted emergent behavior of the agents. They also show that rather standard genetic operators (on the team level and the agent level) are already sufficient to find this unwanted behavior.

On cooperation between evolutionary algorithms and other search paradigms

We present a multi-agent based approach for achieving cooperation between search systems employing different search paradigms. The search agents periodically interrupt their search, select interesting information from their states that is transmitted to the other agents, filter the information sent to them with respect to their own demands, integrate the remaining information into their search, and then continue the search. There are different kinds of information to be exchanged and the selection is both success- and demand-driven. We demonstrate the usefulness of this approach by coupling a search system based on a genetic algorithm and a branch-and-bound based system for job-shop-scheduling. Our experiments show that the cooperation results in finding better solutions within a given time limit and in finding solutions comparable to those generated by the best system working alone in less time. The speed-up factors for some examples even exceed the number of agents (computers) used.

DISCOUNT - A Distributed and Learning Equational Prover

The DISCOUNT system is a distributed equational theorem prover based on the teamwork method for knowledge-based distribution. It uses an extended version of unfailing Knuth–Bendix completion that is able to deal with arbitrarily quantified goals. DISCOUNT features many different control strategies that cooperate using the teamwork approach. Competition between multiple strategies, combined with reactive planning, results in an adaptation of the whole system to given problems, and thus in a very high degree of independence from user interaction. Teamwork also provides a suitable framework for the use of control strategies based on learning from previous proof experiences. One of these strategies forms the core of the expert global_learn, which is capable of learning from successful proofs of several problems. This expert, running sequentially, was one of the entrants in the competition (DISCOUNT/GL), while a distributed DISCOUNT system running on two workstations was another en trant.

Semi-automating small-scale source code reuse via structural correspondence

Developers perform small-scale reuse tasks to save time and to increase the quality of their code, but due to their small scale, the costs of such tasks can quickly outweigh their benefits. Existing approaches focus on locating source code for reuse but do not support the integration of the located code within the developers system, thereby leaving the developer with the burden of performing integration manually. This paper presents an approach that uses the developers context to help integrate the reused source code into the developers own source code. The approach approximates a theoretical framework (higher-order anti-unification modulo theories), known to be undecidable in general, to determine candidate correspondences between the source code to be reused and the developers current (incomplete) system. This approach has been implemented in a prototype tool, called Jigsaw, that identifies and evaluates candidate correspondences greedily with respect to the highest similarity. Situations involving multiple candidate correspondences with similarities above a defined threshold are presented to the developer for resolution. Two empirical evaluations were conducted: an experiment comparing the quality of Jigsaws results against suspected cases of small-scale reuse in an industrial system; and case studies with two industrial developers to consider its practical usefulness and usability issues.

Ontology-guided learning to improve communication between groups of agents

We present a general method for agents using ontologies as part of their knowledge representation to teach each other concepts to improve their communication and thus cooperation abilities. Our method aims at getting positive and negative examples for a concept only very vaguely understood by a particular agent from the other agents. This agent then uses one of the known concept learning methods to learn the concept in question, involving the other agents again by taking votes in case of conflicts in the received knowledge. This method allows agents that are not sharing common ontologies to establish common grounds on concepts known only to some of them, if these common grounds are needed during cooperation. While the concepts learned by an agent are only compromises between the views of the other agents, the method nevertheless enhances the autonomy of agents using it substantially.

Evolutionary online learning of cooperative behavior with situation-action pairs

We present a concept to use off-line learning approaches to achieve online learning of cooperative behavior of agents and instantiate this concept for evolutionary learning with agents based on prototype situation-action-pairs and the nearest-neighbor rule. For such an agent model also modeling of other agents can be achieved using the agents own architecture with situation-action-pairs derived from observations. We tested our online learning agents for different variants of the pursuit game and characterize the aspects of variants for which our online learning agents outperform off-line learning ones. Since our concept also allows a smooth transition from off-line learning to online learning and vice versa, the resulting system is able to win much more game variants than systems using either on- or off-line learning exclusively.

Distributing Equational Theorem Proving

In this paper we show that distributing the theorem proving task to several experts is a promising idea. We describe the team work method which allows the experts to compete for a while and then to cooperate. In the cooperation phase the best results derived in the competition phase are collected and the less important results are forgotten. We describe some useful experts and explain in detail how they work together. We establish fairness criteria and so prove the distributed system to be both complete and correct. We have implemented our system and show by non-trivial examples that drastical time speed-ups are possible for a cooperating team of experts compared to the time needed by the best expert in the team.

Evolutionary behavior testing of commercial computer games

We present an approach to use evolutionary learning of behavior to improve testing of commercial computer games. After identifying unwanted results or behavior of the game, we propose to develop measures on how near a sequence of game states comes to the unwanted behavior and to use these measures within the fitness function of a GA working on action sequences. This allows to find action sequences that produce the unwanted behavior, if they exist. Our experimental evaluation of the method with the FIFA-99 game and scoring a goal as unwanted behavior shows that the method is able to find such action sequences, allowing for an easy reproduction of critical situations and improvements to the tested game.

Determining detailed structural correspondence for generalization tasks

Generalization tasks are important for continual improvement to the design of an evolving code base, eliminating redundancy where it has accumulated. An important step in generalization is identifying the detailed structural correspondence between two pieces of code being considered for generalization. Unfortunately, tool support for this step is insufficient, leaving the developer to resort to tedious and error-prone manual determination of correspondence. This paper presents an approach for automatically determining correspondences as an early step in a generalization task. The approach is implemented in a proof-of-concept plug-in to the Eclipse integrated development environment. Two small empirical evaluations of the tool have been conducted: a comparison between human attempts to determine detailed correspondences and those of the tool; and, a comparison of the use of the tool to the use of diff/CCFinder in performing generalization tasks.

Learning domain knowledge to improve theorem proving

We present two learning inference control heuristics for equational deduction. Based on data about facts that contributed to previous proofs, evaluation functions learn to select equations that are likely to be of use in new situations. The first evaluation function works by symbolic retrieval of generalized patterns from a knowledge base, the second function compiles the knowledge into abstract term evaluation trees. We analyze the performance of the two heuristics on a set of examples and demonstrate their usefulness. We also show that these strategies are well suited for cooperation in the framework of the knowledge based distribution method teamwork.