Procedural content generation via Machine Learning (PCGML) is the umbrella term for approaches that generate content for games via machine learning. One of the benefits of PCGML is that, unlike search or grammar-based PCG, it does not require hand authoring of initial content or rules. Instead, PCGML relies on existing content and black box models, which can be difficult to tune or tweak without expert knowledge. This is especially problematic when a human designer needs to understand how to manipulate their data or models to achieve desired results. We present an approach to Explainable PCGML via Design Patterns in which the design patterns act as a vocabulary and mode of interaction between user and model. We demonstrate that our technique outperforms non-explainable versions of our system in interactions with five expert designers, four of whom lack any machine learning expertise.
Procedural Content Generation (PCG), represents a field of research into, and a set of techniques for, generating game content algorithmically. PCG historically requires a significant amount of human-authored knowledge to generate content, such as rules, heuristics, and individual components, creating a time and design expertise burden. Procedural Content Generation via Machine Learning (PCGML) attempts to solve these issues by applying machine learning to extract this design knowledge from existing corpora of game content (Summerville et al. 2017). However, this approach has its own weaknesses; Applied naively, these models require machine learning literacy to understand and debug. Machine learning literacy is uncommon, especially among those designers who might most benefit from PCGML.
Explainable AI represents a field of research into opening
up black box Artificial Intelligence and Machine Learning
models to users
Design patterns
We present an underlying system for a potential cocreative PCGML tool, intended for designers without ML expertise. This system takes user-defined design patterns for a target level, and outputs a PCGML model. The design patterns provided by designers and generated by our system can be understood as labels on level structure, which allow our PCGML model to better represent and reflect the design values of an individual designer. This system has two major components: (1) a classification system that learns to classify level structures with the user-specified design pattern labels. This system ensures a user does not have to label all existing content to train (2) a level generation system that incorporates the user’s level design patterns, and can use these patterns as a vocabulary with which to interact with the user. For example, generating labels on level structure to represent the model’s interpretation of that structure to the user.
The rest of this paper is organized as follows. First, we relate our work to prior, related work. Second, we describe our Explainable PCGML (XPCGML) system in terms of the two major components. Third, we discuss the three evaluations we ran with five expert designers. We end with a discussion of the systems limitations, future work, and conclusions. Our major contributions are the first application of explainable AI to PCGML, the use of a random forest classifier to minimize user effort, and the results of our evaluations. Our results demonstrate both the promise of these pattern labels “intro”
… in improving user interaction and a positive impact on the underlying model’s performance.
There exist many prior approaches to co-creative or
mixedinitiative design agents and editors
Procedural content generation via Machine Learning (Summerville et al. 2017) is a relatively new field, focused on generating content through machine learning methods. The majority of PCGML approaches represent black box methods, without any prior approach focused on explainability or co-creativity. We note some discussion in the Summerville et al. survey paper on potential collaborative approaches. Summerville (2016a) explored adapting levels to players, but no work to our knowledge looks at adapting models to individual designers.
Super Mario Bros. (SMB) represents a common area
of research into PCGML (Dahlskog and Togelius 2012;
Summerville and Mateas
Explainable AI represents an emerging field of research
Design patterns represent a well-researched approach to
game design
The approach presented in this paper builds an Explainable
PCGML model based on existing level structure and an
expert labeling design patterns upon that structure. We chose
Super Mario Bros. as a domain given its familiarity to the
game designers who took part in our evaluation. The
general process for building a final model is as follows: First,
users label existing game levels with the game design
patterns they want to use for communicating with the system.
For example, one might label both areas with large amounts
of enemies and areas that require precise jumps as
“challenges”. The exact label can be anything as long as it is used
consistently. Given this initial user labeling of level
structure, we train a random forest classifier to classify additional
level structure according to the labeled level chunks
We make use of Super Mario Bros. as our domain, and, in
particular, we utilize those Super Mario Bros levels present
in the Video Game Level Corpus
In total we identified thirty unique classes of level components, and make use of a matrix representation for each level section of size 8 8 30. The first two dimensions determine the tiles in the x and y axes, while the last dimension represents a one-hot vector of length 30 expressing component class. This vector is all 0’s for any empty tile of a Super Mario Bros. level, and otherwise has 1’s at the index associated with that particular level component. Thus, we can represent all level components, including background decoration. We note that we treat palette swaps of the same component as equivalent in class.
We make use of the SciPy random forest classifier
Our goal for the design pattern label classifier is to minimize
the amount of work and time costs for a potential user of the
system. Users have to label level structure with the design
patterns they would like to use, but the label classifier
ensures they do not have to hand-label all available levels. The
classifier for this task must be able to perform given access
to whatever small amount of training data a designer is
willing to label for it, along with being able to easily update its
model given potential feedback from a user. We anticipate
the exact amount of training data the system has access to
will differ widely between users, but we do not wish to
overburden authors with long data labeling tasks. Random forest
classifiers are known to perform reasonably under these
constraints
The random forest model takes in an eight by eight level section and returns a level design pattern (either a userdefined design pattern or none). We train the random forest model based on the design pattern labels submitted by the user. We use a forest of size 100 with a maximum depth of 100 in order to encourage generality.
In the case of an interactive, iterative system the random forest can be easily retrained. In the case where the random forest classifier correctly classifies any new design pattern there is no need for retraining. Otherwise, we can delete a subset of the trees of the random forest that incorrectly classified the design pattern, and retrain an appropriate number of trees to return to the maximum forest size on the existing labels and any additional new information.
Even with the design pattern level classifier this system requires the somewhat unusual step of labeling existing level structure with design patterns a user finds important. However, this is a necessary step for the benefit of a shared vocabulary, and labeling content is much easier than designing new content. Further, we note that when two humans collaborate they must negotiate a shared vocabulary.
The existing level generation system is based on an autoencoder, and we visualize its architecture in Figure 1. The input comes in the form of a chunk of level content and the associated design patterns label, such as “intro” in the figure. This chunk is represented as an eight by eight by thirty input tensor plus a tensor of size n where n indicates the total number of design pattern labels given by the user. This last vector of size n is a one-hot encoded vector of level design pattern labels.
After input, the level structure and design pattern label
vector are separated. The level structure passes through a
two layer convolutional neural network (CNN). We note that
we placed a dropout layer in between the two CNN layers to
allow better generalization. After the CNN layers the output
of this section and the design patterns vector recombine and
pass through a fully connected layer with relu activation to
an embedded vector of size 512. We note that, while large,
this is much smaller than the 1920+n features of the input
layer. The decoder section is an inverse of the encoder
section of the architecture, starting with a relu fully connected
layer, then deconvolutional neural network layers with
upsampling handling the level structure. We implemented this
model with an adam optimizer and mean square loss. Note
that for the purposes of evaluation this is a standard
autoencoder. We intend to make use of a variational autoencoder in
future work
Our system has two major parts: (1) a random forest classifier that attempts to label additional content with userprovided design patterns to learn the designer’s vocabulary and (2) an autoencoder over level structure and associated patterns for generation. In this section we present three evaluations of our system. The first addresses the random forest classifier of labels, the second the entirety of the system, and the third addresses the limiting factor of time in human computer interactions. For all three evaluations we make use of a dataset of levels from Super Mario Bros. labeled by five expert designers.
We reached out to ten design experts to label three or more
Super Mario Bros. levels of their choice to serve as a dataset
for this evaluation. We do not include prior, published
academic patterns of Super Mario Bros. levels (e.g.
Adam Le Doux: Le Doux is a game developer and designer best known for his Bitsy game engine. He is currently a Narrative Tool Developer at Bungie.
Dee Del Rosario: Del Rosario is an events and community organizer in games with organizations such as Different Games Collective and Seattle Indies, along with being a gamedev hobbyist. They currently work as a web developer and educator.
Kartik Kini: Kini is an indie game developer through his studio Finite Reflection, and an associate producer at Cartoon Network Games.
Gillian Smith: Smith is an Assistant Professor at WPI. She focuses on game design, AI, craft, and generative design.
Kelly Snyder: Snyder is an Art Producer at Bethesda and previously a Technical Producer at Bungie.
All five of these experts were asked to label their choice of three levels with labels that established “a common language/vocabulary that you’d use if you were designing levels like this with another human”. Of these experts only Smith had any knowledge of the underlying system. She produced two sets of design patterns for the levels she labeled, one including only those patterns she felt the system could understand and the second including all patterns that matched the above criteria. We refer to these labels as Smith and Smith-Naive through the rest of this section, respectively.
These experts labeled static images of non-boss and
nonunderwater Super Mario Bros. levels present in the Video
Game Level Corpus (VGLC)
We include some summarizing information about these six sets of design pattern labels in Table 1. Specifically, we include the total number of labels and the top three labels, sorted by frequency and alphabetically, of each set. Each expert produced very distinct labels, with less than one percent of labels shared between different experts. We include the first example for the top label for each set of design patterns in Figure 2. Even in the case of Kini and Del Rosario, where there is a similar area and design pattern label, the focus differs. We train six separate models, one for each set of design pattern labels (Smith has two).
In this section we seek to understand how well our random
forest classifier is able to identify design patterns in level
structure. For the purposes of this evaluation we made use
of AlexNet as a baseline
Our major focus is training and test accuracy across the folds. We summarize the results of this evaluation in Table 2, giving the average training and test accuracies across all folds along with the standard deviation. We note that in all instances our random forest (RF) approach outperformed AlexNet CNN in terms of training accuracy, and nearly always in terms of test accuracy. We note that given more training time AlexNet’s training accuracy might improve, but at the cost of test accuracy. We further note that AlexNet was on average one and a half times slower than the random forest in terms of training time. These results indicate that our random forest produces a more general classifier compared to AlexNet.
We note that our random forest performed fairly consistently in terms of training accuracy, at around 85%, but that the test accuracy varied significantly. Notably, the test accuracy did not vary according to the the number of training samples or number of labels per expert. This indicates that individual experts identify patterns that are more or less easy to classify automatically. Further we note that Snyder and Del Rosario had very low testing error across the board, which indicates a large amount of variance between tagged examples. Despite this, we demonstrate the utility of this approach in the next section.
We hypothesize that the inclusion of design pattern labels into our autoencoder network will improve its overall representative quality. Further, that the use of an automatic label classifier will allow us to gather sufficient training data to train the autoencoder. This evaluation addresses both these hypotheses. We draw upon the same dataset and the same three folds from the prior evaluation and create three variations of our system. The first autoencoder variation has no design pattern labels and is trained on all 8 8 chunks of level instead of only those chunks labeled or autolabeled with a design pattern. Given that this means fewer features and smaller input and output tensors, this model should outperform our full model unless the design pattern labels improve overall representative quality. The second autoencoder variation does not make use of the automatic design pattern label classifier, thus greatly reducing the training data. The last variation is simply our full system. For all approaches we trained till training error converged. We note that we trained a single ’no labels’ variation and tested it on each expert, but trained models for the no automatic classifier and full versions of our approach for each expert.
Given these three variations, we chose to measure the difference in structure when the autoencoder was fed the test portions of each of the three folds. Specifically we capture the number of incorrect structure features predicted. This can be understood as a stand in for representation quality, given that the output of the autoencoder for the test sample will be the closest thing the autoencoder can represent to the test sample.
We give the average number and standard deviation of incorrect structural features/tiles over all three folds in Table 2. We note that the minimum value here would be 0 errors and the maximum value would be 8 8 30 or 1920 incorrect structural feature values. For every expert except for Kini, who authored the smallest number of labels, our full system outperformed both variations. While some of these numbers are fairly close between the full and no labels variation, the values in bold were significantly lower according to the paired Wilcoxon Mann Whitney U test (p < 0:001).
Given the results in Table 3. We argue that both our hypotheses were shown to be correct, granted that the expert gives sufficient labels, with the cut-off appearing to be between Kini’s 28 and Del Rosario’s 38. Specifically the representation quality is improved when labels are used, and the label classifier improves performance over not applying the label classifier.
A major concern for any co-creative tool based on Machine Learning is training time. In the prior autoencoder evaluation, both the no labels and full versions of our system took hours to train to convergence. This represents a major weakness, given that in some co-creative contexts designers may not want to wait for an offline training process, especially when we anticipate authors wanting to rapidly update their set of labels. Given these concerns, we evaluate a variation of our approach utilizing transfer learning. This drastically speeds up training time by adapting the weights of a pretrained network on one task to a new task.
We make use of student-teacher or born again neural networks, a transfer learning approach in which the weights of a pre-trained neural network are copied into another network of a different size. In this case we take the weights from our no labels autoencoder from the prior evaluation, copy them into our full architecture, and train from there. We construct two variations of this approach, once again depending on the use of the random forest label classifier or not. We compare both variations to the full and no labels system from the prior evaluation, using the same metric.
We present the results of this evaluation in Table 4. We note that, while the best performing variation did not change from the prior variation, in all cases except for the Kini models, the transfer approaches got closer to the full variation approach, sometimes off by as little as a fraction of one structure feature. Further, these approaches were significantly faster to train, with the no automatic labeling transfer approach training in an average of 4.48 seconds and the automatic labeler transfer approach training in an average of 144.92 seconds, compared to the average of roughly five hours of the full approach on the same computer. This points to a clear breakdown in when it makes sense to apply what variation of our approach, depending on time requirements and processing power. In addition, it continues to support our hypotheses concerning the use of automatic labeler and personal level design pattern labels.
We do not present a front-end or interaction paradigm for the use of this Explainable PCGML system, as we feel such implementation details will depend upon the intended audience. However, it is illustrative to give an example of how the system could be used. In Figure 3 we present an example of the two training examples of the pattern “completionist reward” labeled by the expert Dee Del Rosario. The full system, including the random forest classifier, trains on these examples (and the other labels from Del Rosario), and is then given as input the eight by eight chunk with only the floating bar within it on the left of the image along with the desired label “completionist reward”. One can imagine that Del Rosario as a user wants to add a reward to this section, but doesn’t have any strong ideas. Given this input the system outputs the image on the right.
We asked Del Rosario what they thought of the
performance of the system and whether they considered this
output matched their definition of completionist reward. They
replied “Yes – I think? I would because I’m focusing on
the position of the coins.” We note that Del Rosario did not
see the most decisive patch when making this statement,
which we extracted as in
In this paper, we present an approach to explainable PCGML (XPCGML) through user-authored design pattern labels over existing level structure. We evaluate our autoencoder and random forest labeler components on levels labeled by game design experts. These labels serve as a shared language between the user and level design agent, which allows for the possibility of explainability and meaningful collaborative interaction. We intend to take our system and incorporate it into a co-creative tool for novice and expert level designers. To the best of our knowledge this represents the first approach to explainable PCGML.
We want to thank our five expert volunteers. This material is based upon work supported by the National Science Foundation under Grant No. IIS-1525967. We would also like to thank the organizers and attendees of Dagstuhl Seminar 17471 on Artificial and Computational Intelligence in Games: AI-Driven Game Design, where the discussion that lead to this research began. Abadi, M.; Barham, P.; Chen, J.; Chen, Z.; Davis, A.; Dean, J.; Devin, M.; Ghemawat, S.; Irving, G.; Isard, M.; et al.