The thesis introduces a new approach that utilises plan traces, which represent plans executed by human players. A set of input plan traces is used for constructing an abstraction model generalising a planning domain in a video game. The model has the form of hierarchically nesting regions that partition the state space. Regions of the state space are defined implicitly, which allows identifying sets of states without explicitly specifying and storing them in the memory. Such a hierarchical structure can be applied to estimating the distance between any pair of states in a state-space graph. In practice, the model is prepared before planning is performed. It is employed by the heuristic to accelerate the process of solving planning problems that dynamically appear during the game.

Bartłomiej Józef Dzieńkowski : Classical Planning Supported by Plan Traces for Video Games. Doctoral Dissertation, Wroclaw University of Science and Technology, Faculty of Computer Science and Management, Department of Computational Intelligence, 2018.

Game theory is a tool that may be used to model a player as an intelligent being – one who seeks to optimise his own performance while taking into account the performance of his opponent. However, it is often challenging to apply the theory in practice. In the naval environment, this approach may be used, for instance, to find the best strategy for an Autonomous Underwater Vehicle (AUV) while considering the intelligence of the submarine opponent. Classic approaches based on Minimax suffer from an explosion of states, and they are difficult to use in real-time. The paper introduces an approach that improves the Minimax algorithm in a complex naval environment. It assumes limited and scalable computational resources. The approach takes advantage of a flexible utility function based on a neural network with parameters tuned by a genetic algorithm.

Dzieńkowski, B., Strode, C., Markowska-Kaczmar, U.: Employing Game Theory and Computational Intelligence to Find the Optimal Strategy of an Autonomous Underwater Vehicle against a Submarine. Proceedings of the 2016 Federated Conference on Computer Science and Information Systems, ACSIS, Vol. 8, pp. 31–40, 2016.

The paper presents a new method of building a hierarchical model of the state space. The model is extracted fully automatically from game replays that store executed plan traces. It is used by a novel approach for estimating the distance between states in a state-space graph. The estimate is applied in the A* algorithm as a heuristic function to reduce the search space. The method was validated using the game Smart Blocks. It is a testbed environment for studying methods that benefit from game replay analysis. The proposed heuristic is dedicated to difficult classical planning problems, for which problem-specific or automated heuristics are difficult to obtain.

Dzieńkowski, B., Markowska-Kaczmar, U.: A* Heuristic Based on a Hierarchical Space Model Extracted from Game Replays. Proceedings of the 2016 Federated Conference on Computer Science and Information Systems, ACSIS, Vol. 8, pp. 21–30, 2016.

Plan reconstruction is a task that appears in plan recognition and learning by practice. The document introduces a method of building a goal graph, which holds a reconstructed plan structure, based on analysis of data provided by the observation of an agent. The approach is STRIPS-free and uses only a little knowledge about an environment. The process is automatic; thus, it does not require manual annotation of observations. The paper provides details of the developed algorithm. In the experimental study, properties of a goal graph were evaluated. Possible application areas of the method are described.

Dzieńkowski, B., Markowska-Kaczmar, U.: Plan and Goal Structure Reconstruction: an Automated and Incremental Method Based on Observation of a Single Agent. 11th IFIP TC 8 International Conference, CISIM 2012, LNCS, Vol. 7564, pp. 290-301, 2012.

The paper presents a survey of multi-agent systems (MAS) with adversarial teams competing in a dynamic environment. Agents within teams work together against an opposing group of agents to fulfil their contrary goals. The article introduces the specificity of such an environment and indicates fields of cooperation. It emphasises the role of opponent analysis. Popular planning and learning methods are considered. Next, possible fields of practical application are discussed. The final part of the paper presents a summary of applied machine learning methods and points up future development directions.

Dzieńkowski, B., Markowska-Kaczmar, U.: Agent Cooperation within Adversarial Teams in Dynamic Environment – Key Issues and Development Trends. Transactions on Computational Collective Intelligence VI, LNCS, Vol. 7190, pp. 146-169, 2012.

The paper discusses a problem of inter-rater agreement for the case of human observers who judge how similar pairs of images are. In such a case, significant differences in judgement appear among the group of people. We have observed that for some pairs of images all values of similarity ratings are assigned by various people with approximately the same probability. To investigate this phenomenon in a more thorough manner we performed experiments in which inter-rater coefficients were used to measure the level of agreement for each given pair of images and for each pair of human judges. The results obtained in the experiments suggest that the variation of the level of agreement is considerable among pairs of images as well as among pairs of people.

Michalak, K., Dzieńkowski, B., Hudyma, E., Stanek, M.: Analysis of Inter-rater Agreement among Human Observers Who Judge Image Similarity. Advances in Intelligent and Soft Computing, ISSN 1867-5662, Vol. 95, pp. 249-258, 2011.

The paper presents the architecture of a biologically inspired agent. The agent’s movement is directly controlled by Spiking Neural Network. The neural network is trained by a genetic algorithm. The agents move in a 3D physical environment. Their primary purpose is to translocate themselves using a virtual body structure and muscles. This approach is inspired by a biological assumption, where the neural network receives signals from sensors and directly controls the muscles.

Dzieńkowski, B., Markowska-Kaczmar, U.: Biologically Inspired Agent System Based on Spiking Neural Network. Proceedings of the 4th KES International Conference on Agent and Multi-agent Systems: Technologies and Applications, LNCS, Vol. 6071, pp. 110-119, 2010.

Dzieńkowski, B., Kucharski, K., Chowaniec, Ł., Zyśk, D., Słowiński, P., Węgrzyn, B., Markowska-Kaczmar, U.: Artificial Life (Sztuczne życie). In: Markowska-Kaczmar, U., Kwaśnicka, H. (Eds.): Nature-inspired Metaheuristics in Applications (Metaheurystyki inspirowane naturą w zastosowaniach). Oficyna Wydawnicza Politechniki Wrocławskiej, Wrocław, pp. 25-48, 2013.