Dimitrios Kalles

GA Tree: genetically evolved decision trees

We use genetic algorithms to evolve classification decision trees. The performance of the system is measured on a set of standard discretized concept learning problems and compared (very favorably) with the performance of two known algorithms (C4.5, OneR).

On verifying game designs and playing strategies using reinforcement learning

In this paper we elaborate on the application of reinforcement learning to the design of a new strategy game. We deal with playability and learning issues, attempting to use intelligently generated self-playing sequences to determine playability of various initial board configurations. The machines a priori knowledge about the game is restricted to the rules only, so, the initially encouraging and intuitive results suggest that this design verification strategy may be useful to a board range of design problems.

Students working for students on programming courses

In this paper we offer a report on a university-level programming laboratory course that has been designed on top of a programming library. The course enforces soft skills, such as code inspection and team working, sharpens implementation skills and creates a bridge between introductory, language-specific instruction and senior-year full-blown programming projects that are usually large but not attending to soft skills. Quite as importantly, it has also delivered a working research tool.

Interactive verification of game design and playing strategies

Reinforcement learning is considered as one of the most suitable and prominent methods for solving game problems due to its capability to discover good strategies by extended se self-training and limited initial knowledge. In this paper we elaborate on using reinforcement learning for verifying game designs and playing strategies. Specifically, we examine a new strategy game that has been trained on self-playing games and analyze the game performance after human interaction. We demonstrate, through selected game instances, the impact of human interference to the learning process, and eventually the game design.

Decision Tree Toolkit: A Component-Based Library of Decision Tree Algorithms

This paper reports on the development of a library of decision tree algorithms in Java. The basic model of a decision tree algorithm is presented and then used to justify the design choices and system architecture issues. The library has been designed for flexibility and adaptability. Its basic goal was an open system that could easily embody parts of different conventional as well as new algorithms, without the need of knowing the inner organization of the system in detail. The system has an integrated interface (ClassExplorer), which is used for controlling and combining components that comprise decision trees. The ClassExplorer can create objects on the fly, from classes unknown during compilation time. Conclusions and considerations about extensions towards a more visual system are also described.

STABLE DECISION TREES: USING LOCAL ANARCHY FOR EFFICIENT INCREMENTAL LEARNING

This work deals with stability in incremental induction of decision trees. Stability problems arise when an induction algorithm must revise a decision tree very often and oscillations between similar concepts decrease learning speed. We introduce a heuristic and an algorithm with theoretical and experimental backing to tackle this problem.

Evolving Multi-Layer Neural Networks for Othello

Othello has long been a favorite AI subject due to its very simple rules, its very low branching factor, its well defined strategic concepts and its dramatic changes in board topology as a game unfolds. In this paper, we investigate several neural network architectures using co-evolutionary learning techniques, with the objective to learn to play Othello strongly. Our resulting neural networks were able to learn to play the game at an expert-master level and to discover advanced strategies, within a few thousand generations, without any prior knowledge, beyond the game rules.

A Pendulum Effect of Expert Playing in Games

When learning how to play a strategy board game, one can measure the relative effectiveness of the learned policies by assessing how often a player wins and how easily these wins are scored. Experimental evidence also shows that when one of the competing players is trained by a sophisticated tutor, performance benefits also flow to the opponent. We present comprehensive experimental evidence that the level of tutor effectiveness is best demonstrated by the improvement of the tutored players opponent, this performance change is termed the pendulum effect.

Controlled flux results in stable decision trees

This work deals with stability in incremental induction of decision trees. Stability problems arise when an induction algorithm must revise a decision tree very often and oscillations between similar concepts decrease learning speed. We introduce a heuristic and an algorithm with theoretical and experimental backing to tackle this problem.