Thesis

The inspiration for this work comes from the belief that computers can do more than perform computational tasks, but rather tackle a concept that most people think is impossible for a computer to achieve, due to being unique to the human being: Creativity. To demonstrate this, the proposed solution consists of a level design tool with co-creativity at its core, this is achieved by presenting the user with suggestions of possible modifications to their current level. There are two types of suggestions: level layout and lighting. Level layout consists on which tiles a player can walk through or not. And the latter, lighting is how each part of the level is illuminated to provide the player with the best experience possible. These suggestions are generated by genetic algorithms with parameters that can be adjusted through an user-intuitive interface, either it being more evolved towards level layout or the lighting setup.

Machine Learning is now an exciting field due to the increased computing power available today. Realworld application discoveries are increasing, but also new algorithms and strategies developed in this area. Games have always been a great playground for studying these new ideas and Geometry Friends is a great example. It is a multi-player puzzle game with its own competition that aims to distinguish the best cooperative and non-cooperative agents. The single player problem has had many different approaches over the years with very satisfactory results. Attention now turns to the resolution of the cooperation component associated with multi-player gameplay. In this paper, we propose a consistent multi-agent learning system architecture, inspired by the single agent success of using a weighted directed graph and Reinforcement Learning. The motivation is to build a solid foundation for future cooperation solutions that want to expand and exploit Machine Learning knowledge. The final results will demonstrate that our system outperforms all proposals submitted to date, based on a relative simple structure giving the complex demands of a multi-agent environment.

Earthquakes continue to be one of the most destructive natural disasters. Even with the technological advances that have been made, it is impossible to predict when the next earthquake will occur or what magnitude it will have, so it is critical that everyone is prepared for an earthquake. In 2014 a serious game was developed to teach people, particularly primary school children, which behaviours should be adopted before, during and after an earthquake. This game, Treme-Treme, had a huge success and 5 years later still continues to be used by elementary school teachers as a complement to teaching. In addition, it was also adopted by schools of children with educational difficulties, having a greater impact than initially thought. The opportunity has now arisen to continue the development of this game, and this work portrays the whole process inherent to the realisation of the new version. Over the years, Treme-Treme's code has become obsolete, so the game has been successfully rebuilt from scratch and improved, using a new platform, Godot. Game design was rethought and successfully implemented to provide an improved player experience. The results got with the tests show that Treme-Treme's new solution provides better engagement with children than the previous one, transmitting the intended pedagogical knowledge more effectively and representing a better tool for teachers.

In game design, one of the most important tasks is associated with the playtesting process, as this is where game designers are able to understand if the game experience they are trying to create is indeed being passed to the players. How a player perceives and reacts to a game should be important in the game design process. Work developed in the Deep Learning research field has proved to be a great source of information as to understand how an agent is capable of achieving great runs and scores playing various games from scratch. The capability of testing different games using a screen capture artificial player powered by Convolutional Neural Network allows for a good understanding of how an agent is capable of extracting important features from the game screen without additional information. This Thesis, took the work developed in the Deep Reinforcement Learning field applied to Atari environments, mainly Deep-Q Networks, modulated different types of player limitations, tested and documented the results achieved. The results seem to indicate the existence of different types of playing patterns for different limitations.