DOOM Level Generation using Generative Adversarial Networks (1804.09154v1)

Abstract: We applied Generative Adversarial Networks (GANs) to learn a model of DOOM levels from human-designed content. Initially, we analysed the levels and extracted several topological features. Then, for each level, we extracted a set of images identifying the occupied area, the height map, the walls, and the position of game objects. We trained two GANs: one using plain level images, one using both the images and some of the features extracted during the preliminary analysis. We used the two networks to generate new levels and compared the results to assess whether the network trained using also the topological features could generate levels more similar to human-designed ones. Our results show that GANs can capture intrinsic structure of DOOM levels and appears to be a promising approach to level generation in first person shooter games.

• The paper demonstrates that incorporating topological features in conditional GANs improves generated DOOM levels compared to unconditional models.
• The study transforms levels into images capturing walls, walkable areas, and items, effectively training GANs to replicate design patterns.
• Results show that GAN-generated levels improve in visual and structural fidelity over time, promising automation in game design.

DOOM Level Generation using Generative Adversarial Networks

The paper "DOOM Level Generation using Generative Adversarial Networks" investigates the use of GANs, a class of neural networks known for generating data with patterns similar to a given dataset, to model and create new levels for the iconic first-person shooter game, DOOM. This paper contributes to procedural content generation (PCG) in video games, emphasizing reducing content creation time and effort by automating level design.

Research and Methods

The authors focused on two primary models of GANs for level generation. The first, an unconditional GAN, used a dataset composed solely of images derived from existing DOOM levels. The second, a conditional GAN, extended this approach by including topological features extracted from the level design, such as area size and room configurations, as additional inputs. They aimed to determine whether these additional features could enhance the generation capabilities of GANs.

Each DOOM level was transformed into a set of images representing significant elements like walkable areas, walls, and items, allowing the GANs to learn from and replicate these features. The research leveraged a selection from the idgames archive, a vast repository of user-generated DOOM levels, focusing on simple topologies that fit computational constraints.

Results

The numerical analysis revealed that both GAN models succeeded in generating levels with characteristics similar to their human-designed counterparts. However, the conditional GAN showed a slightly superior performance in structural resemblance and visual quality due to the integration of additional level features. The training process demonstrated that the GANs were capable of enhancing level aesthetics over time, as observed through metrics like the Structural Similarity Index (SSIM) and other proprietary evaluation metrics related to visual and structural fidelity.

One crucial finding is the conditional GAN's capacity to utilize structured input features to produce levels that better emulate human design, suggesting their potential for more sophisticated procedural generation tasks in the gaming industry. An intriguing aspect of the paper was the ability to measure generated levels' complexity, which could guide future work in tailoring specific game dynamics through level design.

Implications and Future Research Directions

The methodology presented in the paper opens pathways for leveraging AI to decrease the time-intensive human effort in game level design. GAN-driven content generation could be instrumental in generating a wide range of game maps, enhancing replayability, and personalizing gaming experiences in real-time. Future research could expand on the scalability of such models, incorporating more complex level designs or applying similar methodologies to other game genres. Bridging the training process with user feedback could also refine the outputs, tailoring them closer to desired gameplay outcomes.

Overall, the application of GANs for DOOM level generation not only showcases the potential of AI in creative domains but also underscores the importance of structured data input in enhancing generative models. This research paves the way for more autonomous and dynamic content creation tools, essential as the video game industry continues to evolve.