Exploring Level Blending Across Platformers via Paths and Affordances

Abstract

Techniques for procedural content generation via machine learning (PCGML) have been shown to be useful for generating novel game content. While used primarily for producing new content in the style of the game domain used for training, recent works have increasingly started to explore methods for discovering and generating content in novel domains via techniques such as level blending and domain transfer. In this paper, we build on these works and introduce a new PCGML approach for producing novel game content spanning multiple domains. We use a new affordance and path vocabulary to encode data from six different platformer games and train variational autoencoders on this data, enabling us to capture the latent level space spanning all the domains and generate new content with varying proportions of the different domains.

Procedural Content Generation via Knowledge Transformation (PCG-KT)

“…Somewhere in the middle, we could have a learned probability table of tile distributions [11]. At the latent end, we could have the weights of a neural network trained on those levels [12] or the latent vector representations of the levels [16], [17].…”

“…We can think of this as equivalent to the amount of shared information between the input and transformed knowledge, regardless of how that information is represented. As an example, Sarkar et al's approach [17] of learning blended latent representations of several input games has a higher content distance than Snodgrass and Sarkar's approach [14] that generates new levels by combining existing parts of existing game levels using binary space partitioning, which in turn has higher content distance than prior PCGML methods that operated on single-game domains.…”

“…1) Approach: This refers to the actual procedure or manner in which the transformation is being performed. This can be through updating a probability table [11], backpropagation through a neural network [12], iteratively growing a graph from existing extracted graphs [18], interpolating between latent vectors [16], [17], or reducing detailed knowledge representations from various domains into a single uniform representation [14], [17], to name a few. For this category, the function used can be nearly any procedure already used in the PCG, ML, creativity, design, or AI literature, including manual and hybrid approaches.…”

“…1) One input domain to one output domain: this is a standard knowledge transfer where knowledge from one domain is transformed to represent knowledge from another domain [9]. 2) Multiple inputs to one output: this is typically a blending approach where the input domains are transformed and combined in some way to get a different output domain (either new or existing) as in [14], [17], [18]. 3) Multiple inputs to multiple outputs: this can also be blending or approximating, but likely in a more exploratory way where the input knowledge is transformed and recombined in various ways to explore different output domains.…”

“…[46] transfer agent policies from Atari Pong to Breakout while [47] perform transfer learning to build agents capable of playing Breakout and Road Fighter. However, note that in each of [43]- [47], while some knowledge Conceptual Blending [13] Hierarchical Bayes Network Levels [16], [17] Variational Autoencoder (VAE) Levels [32] Variational Autoencoder (VAE) Jumps [19], [20] Conditional VAE Levels [33] Graph Structure Knowledge Base Games [34] Manual Conceptual Blending Games Conceptual Expansion [35] Bayesian Networks Levels [7], [18] Probabilistic Graphical Models Games Transfer Learning Domain Adaptation [9] Multidimensional Markov Chains Levels [36], [14] VAE, Binary Space Partitioning Levels…”