Skip to main content

Information Theoretic Evaluation of Random Content Generation in Games

Periodic Reporting for period 2 - INTERCOGAM (Information Theoretic Evaluation of Random Content Generation in Games)

Période du rapport: 2018-12-01 au 2019-11-30

INTERCOGAM’s goal is to develop and use information theory based intrinsic motivation formalisms to evaluate automatically generated game mechanics. Content generation is one of the production bottlenecks of professional game design, which has begun to be addressed by procedural content generation. In this field, search based procedural content generation uses the idea of evolutionary algorithms to represent, modify and adapt games and game content to maximise fun and engagement with the game. One major challenge here is the identification of widely applicable fitness functions, which capture the different aspects of what makes a game fun, such as challenge level, complexity, pacing, etc. INTERCOGAM will relate psychological and game design concepts of game experience to either existing formalisms for intrinsic motivation or develop new ones, where appropriate. Human play testers will then play procedurally generated games and evaluate their own experience, allowing us to verify whether our formalism captures the actual human motivation, and also whether humans indeed act according to certain intrinsic motivations. INTERCOGAM will yield both, a tool to generate new and engaging game ideas, aiding better and faster game design, and provide new insights into how the human mind engages with different worlds where there is no external reward present.
The INTERCOGAM project so far looked at a range of games and game-like frameworks and applied different intrinsic motivation frameworks and other artificial intelligence tools to generate interesting behaviour and make interesting game content.
Particularly noteworthy is the development of the coupled empowerment framework – where a non-player character is imbued with a set of motivations that cause it to act in an antagonistic or helpful way . The strength of this approach is its flexibility and robustness – which makes it applicable to a range of different games without much adaptation. We demonstrated that this framework captures relatively intuitive notions of support or antagonism, and thus provides us with formal building block to build companion or non-player behaviour. We also argued, and showed in simulation, that this approach could be a feasible way to provide some basic ethical framework to robots, giving them basic notion for empowering humans, maintaining their own ability to act and maintaining their operational flexibility.

As part of the development of the coupled empowerment framework we also made several improvements towards more efficient ways of computing empowerment – specifically for discrete and deterministic scenarios. This lower computational load makes it more feasible to apply empowerment to a range of scenarios. It also encourages further development of those tools beyond the proof of concept implementations presented in INTERCOGAM. In particular, there is an increase in the professional game development community to engage and use more modern AI techniques for game testing and design.

The INTERCOGAM project also contributed to the further development of the theory behind intrinsic motivations. We developed the philosophical argument that aims to close the gap between basic assumptions about the mind from an enactive perspective to the development of a mathematical formalism. Furthermore, we also partook in the development of a unifying framework that can express a range of different intrinsic motivation – making it possible to compare these different notions formally and in a simulation. While games offer a great testbed to evaluate and benefit from robust intrinsic motivation models, understanding intrinsic motivation better offers us great insights on how minds operate and how to approach problems in general.
The experiments with human participants towards the end of the project have also produced some of the first empirical data linking human experience and perceptions to different intrinsically motivated behaviours. We investigated the relationship between empowerment and challenge, and produced significant results demonstrating the relationship between perceived warmth and predictive information in a robot human interaction scenario.
INTERCOGAM applied empowerment to complex game scenarios, and was thus able to showcase its robustness beyond previously existing models. This made it viable as a behaviour generation approach, which allowed us to define antagonistic or companion like behaviour on more fundamental principles than before. Extending this further could change the way we generate and think about non-player behaviour generation. Combined with the novel empirical data about human perception of intrinsically motivate behaviour, this demonstrates the feasibility of a novel approach to robust and generic behaviour generation. While many of the approaches and techniques developed in INTERCOGAM still require human evaluation, the existing results are promising, and demonstrate how intrinsically motivated behaviour generation can be used to produce engaging and socially evocative NPCs or robots. This would not only allow for the development of better and more engaging games, but could also provide useful in providing a toolbox for generating robot behaviour. Being able to produce more socially acceptable or socially aware robots would be beneficial to a wide range of HRI applications, including elder care, domestic and service robots, robot therapy, etc.

INTERCOGAM’s other main goal is to provide a metric that can evaluate a range of different games (or experiences) and provide us with a good proxy for the actual experience. Our initial results indicate that an intrinsic motivation-based approach, as advocated by us, could provide us with a metric that actually evaluates the experience of a player with a game (rather than some property of the artefact) across a range of games. This would have major applications, as it would allow game companies to automatically test games without having to hand-craft test for a specific game or scenario.

Our development of the Generative Design for Minecraft competition and the associated framework have also already led to an increased focus on the idea of adaptive procedural content generation. Universities around the world have started to use the GDMC competition as part of their coursework, and this had allowed us to build a community that is interested in various forms of procedural content generation, and their comparison and human evaluation. Noteworthy here is also that the GDMC includes members of the general public as participants and generators of novel ideas, rather than as test subjects – and, as such, its pushes the idea of citizen science further .
An antagonist using different elements in the environment to frustrate a player.
A random test level used to evaluate the UCB Empowerment Algorithm
Christoph Salge in the Game Innovation Lab, NYU
The winning entry of the first AI Settlement Generation Challenge in Minecraft