This paper presents a framework for cooperation between a human and a general game playing agent.... more This paper presents a framework for cooperation between a human and a general game playing agent. Cooperation is defined as two entities causing each other to modify their behaviour to achieve some mutual advantage. Such humancomputer cooperation has the potential to offer insights that can help us improve the performance of artificial agents, as well as improving the performance of humans during certain kinds of strategic interactions. This paper focuses specifically on game playing as a form of strategic interaction. By proposing a framework for cooperation between a human and a general game playing agent, our aim is to create a flexible system that may be applicable to cooperation in other kinds of problem solving and strategic interactions in the future. We evaluate the framework presented in this paper by means of a human study. We observe humans playing games with and without the cooperation of a general game playing agent. We present experimental results of the pilot study as well as proposed changes in the experiment. These changes aim to verify the hypothesis that human-machine cooperation within our framework can indeed lead to mutual advantage.
The 1990s brought the renaissance of mind game programming, leading to several extraordinary achi... more The 1990s brought the renaissance of mind game programming, leading to several extraordinary achievements in machine-human competitions. The best-playing programs surpassed world human champions in chess, checkers, Scrabble, and Othello, and are making a striking progress in Go, poker and bridge. The volume of research projects and research papers devoted to AI and games is too high to even try
Several successful applications of TD-learning methods in various games, which include Samuel’s c... more Several successful applications of TD-learning methods in various games, which include Samuel’s checkers player [277, 279], Tesuaro’s TD-Gammon [317, 318, 319], Baxter et al.’s KnightCap [18, 19], Schraudolph et al.’s Go program [294, 295], or Schaeffer et al.’s TDL-Chinook [288], to name only the best-known examples, confirmed wide applicability of this type of learning in game domain. Quite a lot of research papers devoted to TD-learning and games were published, in which particular attempts are thoroughly described focusing on various aspects of TD paradigm. Based on the amount of the literature one might suppose that all relevant questions concerning this type of learning have already been answered and if there exist any ambiguities they must refer to secondary issues or specific implementation details.
The paper provides an argumentation for potential virtues of developing cognitively-plausible hum... more The paper provides an argumentation for potential virtues of developing cognitively-plausible human-like playing systems, thus advocates a return to the roots of Artificial Intelligence application to games. Such systems are, in particular, expected to be capable of intuitive playing, manifested by efficient search-free move pre-selection and application of shallow-search only during regular move analysis. The main facets of such systems are listed and discussed in the paper. Furthermore, an example of search-free playing system, in the form of a specifically-designed convoluted neural network, is presented to illustrate possible implementation of proposed ideas.
Specialized vs. Multi-game Approaches to AI in Games
In this work, we identify the main problems in which methodology of creating multi-game playing p... more In this work, we identify the main problems in which methodology of creating multi-game playing programs diff ers from single-game playing programs. The multi-game framework chosen in this comparison is General Game Playing, which was proposed at Stanford University in 2005, since it de fines current state-of-the-art trends in the area. Based on the results from the International General Game Playing Competitions and additional results of our agent named MINI-Player we conclude on what defi nes a successful player. The most successful players have been using a minimal knowledge and a mechanism called Monte Carlo TreeSearch, which is simulation-based and self-improving over time.
What are the most important problems of computational intelligence? A sketch of the road to intel... more What are the most important problems of computational intelligence? A sketch of the road to intelligent systems is presented. Several experts have made interesting comments on the most challenging problems.
Uploads
Papers by Jacek Mandziuk