General agents to evaluate automatic generated levels
Game Artificial Intelligence covers different areas, Procedural Content generation (PCG) being one of them, which refers to the generation of game content via automated processes. This content generation process includes declaring a space and list of rules to be able to define and create artifacts, which are generated instances of content. One of the applications of PCG in the game industry is automatic level generation. Agents are entities capable of carrying out certain actions in an environment to achieve a goal. Applying this concept to games, the goal is winning and, therefore, the agent will take a series of decisions to reach the victory. These decisions will be based on its knowledge of the surroundings and the game, so an agent designed to win a certain game will be optimized and developed specifically for that game. For example, an agent developed to play Civilization II, knows it needs to build its civilization, develop it and produce different technologies along the game in order to, either, be the last civilization remained (conquering the other ones), or reach the Alpha Centauri with a spaceship.
General Video Game Playing (GVGP) refers to a game domain where it’s not possible to use knowledge focused on a certain game, as the information is very limited or it does not exist at all. Consequently, the players (or agents) that are used in GVG should be created with algorithms that do not take game-specific knowledge into account. Finding and improving these general algorithms is becoming an important field of research in AI.
When new content is generated by automated processes, the validity of the artifact should be checked to ensure its playability. Common approaches are game-specific and, although they ensure the playability of the generated content, they are limited by the game they are defined for. To solve this limitation, a strong connection between PCG and GVG could be established, as general controllers could be used to validate content generated for any game. My research proposal focuses on the use of GVG agents to evaluate new automated generated levels for games. This evaluation will be answered in two different ways: with a numeric value, and with a quality answer, displayed in natural language, being sensible and readable by a human being. An evaluation tool will be built in order to compare real results given by humans with the results given by the set of agents. The human evaluations could be used as a reference to train the system to be capable to give a sensible answer to any automatic generated level. The response would be displayed in natural language, giving some reasons to the characteristics of the level evaluated, such as difficulty and entertainment. This response is expected to give the motive that would determine if a level should be used or not based on the expectations. For example, after the evaluation, the output would look like “This is a valid level. Also this level would be suitable from beginners as it has been determined that it is easy because (...) Other characteristics to take into consideration are (...)”. Having this kind of response, game designers and developers could easily determine if they would like to include the level in the game or not.
Once a game level evaluator that returns the results as expected is built, it will be investigated how to integrate this analyser in bigger systems that generates levels for different games. The combination of these techniques would allow automatically generating and evaluating levels, being able to accept or discard them based on the results processed by the agents.
Cristina studied a BE in Computer Engineering at Universidad Autónoma de Madrid in Spain. She moved to London in 2013 and worked as web developer for a couple of years. She then decided to change careers opting for the exciting world of Artificial Intelligence and games. She is one of the IGGI students based in University of Essex and her interests include Procedural Content Generation (PCG), General Video Game Playing (GVGP) and Natural Language Processing (NLP). Cris has always enjoyed drawing and painting and though sports were never her strength, she tries and currently plays football. Due to her passion for medieval weapons, she gave fencing a try. As her skills with the sword in real life were unsuccessful, she sticks to video-games now.
Home institution: Essex
Ready to apply?
Once you have identified your potential supervisor, we would encourage you to contact them to discuss your research proposal.Learn More