=Paper=
{{Paper
|id=Vol-2862/paper16
|storemode=property
|title=Conditional Level Generation and Game Blending
|pdfUrl=https://ceur-ws.org/Vol-2862/paper16.pdf
|volume=Vol-2862
|authors=Anurag Sarkar,Zhihan Yang,Seth Cooper
|dblpUrl=https://dblp.org/rec/conf/aiide/SarkarYC20
}}
==Conditional Level Generation and Game Blending==
https://ceur-ws.org/Vol-2862/paper16.pdf
Conditional Level Generation and Game Blending
Anurag Sarkar1 , Zhihan Yang2 and Seth Cooper1
1
Northeastern University
2
Carleton College
sarkar.an@northeastern.edu, yangz2@carleton.edu, se.cooper@northeastern.edu
Abstract
Prior research has shown variational autoencoders (VAEs) to
be useful for generating and blending game levels by learn-
ing latent representations of existing level data. We build on SMB - 〈10011〉 KI - 〈1101〉 MM - 〈10101〉
such models by exploring the level design affordances and
applications enabled by conditional VAEs (CVAEs). CVAEs Figure 1: Example original segments with corresponding el-
augment VAEs by allowing them to be trained using labeled ement labels. Super Mario Bros (SMB)-〈Enemy, Pipe, Coin,
data, thus enabling outputs to be generated conditioned on
Breakable, ?-Mark〉, Kid Icarus (KI)-〈Hazard, Door, Mov-
some input. We studied how increased control in the level
generation process and the ability to produce desired outputs ing Platform, Fixed Platform〉, Mega Man (MM)-〈Hazards,
via training on labeled game level data could build on prior Door, Ladder, Platform, Collectable〉. 0/1 in labels indicate
PCGML methods. Through our results of training CVAEs on absence/presence of corresponding elements in the segment.
levels from Super Mario Bros., Kid Icarus and Mega Man, we
show that such models can assist in level design by generating
levels with desired level elements and patterns as well as pro- of the model to find vectors corresponding to levels with de-
ducing blended levels with desired combinations of games. sired properties. Thus, controllability is achieved via evo-
lutionary search once training has already been performed
and is independent of the model. However, conditional vari-
Introduction ants of both GANs (Mirza and Osindero 2014) and VAEs
Procedural Content Generation via Machine Learning (Sohn, Lee, and Yan 2015) could enable such controllabil-
(PCGML) (Summerville et al. 2018) has emerged as a vi- ity as part of the model itself. These variants allow models
able means of building generative models for game levels to be trained on labeled data and thereby allow generation
by training on levels from existing games. While several to be conditioned on input labels. Thus, when applied for
ML approaches have been utilized for PCG such as LSTMs PCGML, such models could use labels provided by design-
(Summerville and Mateas 2016), Bayes nets (Guzdial and ers to produce controllable level generators without having
Riedl 2016a) and Markov models (Snodgrass and Ontañón to define objective functions and run evolution.
2017), a recent body of work has emerged that focuses on Thus, we train conditional VAEs (CVAEs) on levels from
using latent variable models such as Generative Adversar- Super Mario Bros., Kid Icarus and Mega Man using differ-
ial Networks (GANs) (Goodfellow et al. 2014) and Vari- ent sets of labels corresponding to the presence of various
ational Autoencoders (VAEs) (Kingma and Welling 2013). game elements as well as design patterns. Additionally, we
These models learn latent encodings of the input game lev- train a combined model with levels labeled with the game
els, comprising a continuous latent space which can then be they belong to. Our results show that CVAEs can generate
sampled and explored to generate new levels. Such models levels with and without desired elements and design pat-
have been used both for level generation (Volz et al. 2018; terns, making it a promising model to inform future level
Gutierrez and Schrum 2020) and level blending (Sarkar, design tools. Further, our results suggest that CVAEs can
Yang, and Cooper 2019; Snodgrass and Sarkar 2020). Addi- help perform controllable game blending using different la-
tionally, attempts have been made to make such generation bels provided during generation.
and blending controllable via latent vector evolution (Bon-
trager et al. 2018). This involves optimizing some objective Background
function via evolutionary search in the learned latent space
Controllability has been the focus of much PCG research
Copyright © 2020 for this paper by its authors. Use permitted under with several works developing generators that produce con-
Creative Commons License Attribution 4.0 International (CC BY tent based on designer preferences such as (Liapis, Yan-
4.0). nakakis, and Togelius 2013; Alvarez et al. 2020; Smith
et al. 2009). Besides these systems, other works on con- Mega Man (MM). For training our models, we used 16x16
trollable PCG have included graph-based methods (Valls- level segments cropped out from the levels of each game
Vargas, Zhu, and Ontañón 2017; Dormans 2010) and evolu- by using a sliding window. To enable this, SMB levels and
tionary computation (Togelius et al. 2010). Similarly, a num- horizontal portions of MM levels were padded with 1 and 2
ber of PCGML works have also incorporated controllabil- empty rows respectively. We finally obtained 2643 segments
ity. Snodgrass and Ontañón (2016) used a constraint-based for SMB, 1142 for KI and 2983 for MM. Using these games,
method to control sampling for a multi-dimensional Markov we looked at three different conditioning approaches based
model of Mario levels while Sarkar and Cooper (2018) used on 1) game elements 2) SMB design patterns and 3) game
multiple LSTM models with turn-based weighting to control blending, motivated by wanting to generate levels containing
generation of blended Mario-Kid Icarus levels. In terms of desired elements, exhibiting desired patterns and consisting
latent variable models, Volz et al. (2018) used CMA-ES to of desired combinations of games, respectively. Condition-
evolve vectors in the latent space of a GAN trained on Mario ing is accomplished by associating each input level segment
levels. This produced levels capturing desired characteristics with a corresponding label during the training process. In all
based on the objective used for evolution. A similar method cases, labels take the form of binary-encoded vectors.
was used by Sarkar, Yang, and Cooper (2019) to gener-
ate and blend levels of Mario and Kid Icarus using a VAE Game Elements For game elements, we used a unique set
trained on both games. Recently, Schrum et al. (2020) pro- of conditioning labels for each game. The label vector length
duced a tool enabling user-controlled generation of Mario was determined by the number of different game elements
levels and Zelda-like dungeons by evolving vectors via in- considered for each game with a 0/1 value for a vector el-
teractive exploration of the GAN latent space. In terms of ement indicating the absence/presence of the corresponding
conditional models, Torrado et al. (2019) combined a con- game element in the corresponding level segment. For SMB,
ditional GAN approach with self-attention mechanisms for we considered Enemy, Pipe, Coin, Breakable and Question
controllable generation of GVGAI levels. Our work differs Mark elements, for KI, Hazard, Door, Moving and Station-
in using CVAEs instead of CGANs and not restricting to ary Platforms and for MM, Hazard, Door, Ladder, Plat-
GVGAI. Moreover, their conditioning features were learned forms and Collectables. Thus, we used 5-element binary la-
from input levels and not supplied externally as in our case, bels for SMB and MM and 4-element labels for KI, yield-
which seems more suited to future co-creative applications. ing 25 = 32 unique labels for SMB and MM and 24 = 16
Prior work has demonstrated the utility of conceptual unique labels for KI. Example segments and their corre-
blending (Fauconnier and Turner 1998) for generating new sponding element labels are shown in Figure 1.
levels and even entire games. Gow and Corneli (2015) pre- Design Patterns For SMB design patterns, we picked 10
sented a VGDL-based manual game blending framework to such patterns based on the 23 described by Dahlskog and
create new games by combining elements of existing ones. Togelius (2012). The ones we consider are:
Guzdial and Riedl (2016b) similarly blended Mario levels to • Enemy Horde (EH): group of 2 or more enemies
produce new levels. Sarkar and Cooper (2018) and Sarkar, • Gap (G): 1 or more gaps in the ground
Yang, and Cooper (2019) used LSTMs and VAEs respec- • Pipe Valley (PV): valley created by 2 pipes
tively to blend levels of Mario and Kid Icarus while Sarkar • Gap Valley (GV): valley containing a Gap
et al. (2020) extended the latter to a larger set of games and • Null (empty) Valley (NV): valley with no enemies
incorporated playability into blended levels. In this work, we • Enemy Valley (EV): valley with 1 or more enemies
show that game blending can also be achieved using CVAEs. • Multi-Path (MP): segment split into multiple parts hori-
Conditional VAEs (CVAE) (Sohn, Lee, and Yan 2015; zontally by floating platforms
Yan et al. 2015) allow VAEs (Kingma and Welling 2013) to • Risk-Reward (RR): segment containing a collectable
be conditioned on attributes. Unlike VAEs which are trained guarded by an enemy
in an unsupervised manner, to train CVAEs, each input dat- • Stair Up (SU): ascending stair case pattern
apoint is associated with a label. The encoder learns to use • Stair Down (SD): descending stair case pattern
this label to encode the input into the latent space while the We thus had 10-element binary labels for a total of 210 =
decoder learns to use the label to decode the latent encod- 1024 possible unique labels though a vast majority of these
ing. Thus, such models when trained on game levels could do not occur in the data. For this, we only trained on SMB
allow generation to be conditioned on designer-specified la- levels since we only used SMB design patterns.
bels. Further, the same latent vector can produce different
outputs by varying the conditioning labels. This enables ad- Game Blending For game blending, we trained on seg-
ditional affordances for level design and allows generation to ments from all 3 games taken together with labels indicating
be controlled without having to perform evolutionary search. which game the segments belonged to. We used a 3-element
label with the 1st, 2nd and 3rd elements indicating if the
Method segment was from SMB, KI or MM respectively.
Level Data and Conditioning Generation and Blending using CVAE
We used level data from the Video Game Level Corpus Training CVAEs involves associating each input data in-
(VGLC) (Summerville et al. 2016) for the classic NES plat- stance (in our case, level segments) with a label vector.
formers Super Mario Bros. (SMB), Kid Icarus (KI) and An input instance is concatenated with its label and passed
through the encoder to obtain a latent vector. The same la- 32-Dim 64-Dim 128-Dim
Exact None Exact None Exact None
bel is then concatenated with this latent vector and passed
SMB-Rand 33.5 17.6 32.7 17.1 27.1 19
through the decoder. Thus, the encoder learns to use the pro-
KI-Rand 50.7 10 42 10.5 41.4 9.8
vided labels to learn latent encodings of the data and simi-
MM-Rand 17.4 43.2 15.2 42.5 14.5 43.7
larly, the decoder learns to use the same labels to decode the
SMB-Train 35.1 16.4 33.6 16.4 28.3 17.5
encodings. Since the encoder and decoder use the labels to
KI-Train 49.4 8.3 40.3 9.3 39.5 8.5
learn the encodings between the latent space and the data,
MM-Train 17.8 42.7 15.3 42.1 14.8 42.6
the same latent vector could be made to produce different
outputs by varying the label during generation. These affor-
Table 1: Results of conditioning randomly sampled (‘Rand’)
dances of the conditional VAE could inform level design and
and training (‘Train’) segments using element labels. High-
generation in two specific ways: 1) enable controllable gen-
est Exact values and lowest None values per game are high-
eration by using labels to produce desired content and 2)
lighted in bold.
generate variations of existing content by encoding it into
latent space and decoding it using different labels. Through
this work, we hoped to explore these specific affordances.
For game elements, we trained separate CVAEs for each
game. In all models, both the encoder and decoder consisted
of 4 fully-connected layers with ReLU activation. For SMB
and MM, the conditioning labels were binary vectors of
length 5 while for KI, they were of length 4. Labels for each
input segment were determined by checking for the pres-
ence of the relevant game elements within that segment and
assigning 0 or 1 to the corresponding element of the label.
All models were trained in Pytorch (Paszke et al. 2017) for
10000 epochs using the Adam optimizer with a learning rate
of 0.001 decayed by 0.1 every 2500 epochs.
For design patterns, we trained only on SMB data. Rather
than use all segments obtained by sliding the window across
the levels with redundancy in terms of overlap, we only used
non-overlapping segments. We found non-overlapped seg-
ments better preserve the original design patterns. Since this
led to fewer segments, we additionally used VGLC level
data from Super Mario Bros II: The Lost Levels, to obtain
a total of 407 segments. Labels corresponding to design pat-
terns were assigned manually based on visual inspection. Figure 2: Results of game element conditioning for both
The model architecture was the same as those used for game generated and training levels of SMB (top) with frequencies
elements, but was only trained for 5000 epochs with the of each label in the training data (bottom). X-axis values are
learning rate decay occurring every 1250 epochs. the integer encodings of the equivalent binary label. Results
Finally for blending, we trained on levels from SMB, KI shown for 16 most frequent labels in the training levels.
and MM taken together. Labels indicated the game that the
levels belonged to and were of length 3 with 〈100〉, 〈010〉
and 〈001〉 indicating SMB, KI and MM respectively. Here, Game Elements
we intend to achieve blending by leveraging the fact that it For evaluation, for each game, we randomly sampled 1000
is possible to condition generation using labels that do not latent vectors and conditioned the generation for each vec-
appear in the training set. For instance, though the 3 labels tor using each possible label (32 for SMB and MM, 16 for
above are the only ones used in the training set, we could KI). We then tested if the generated segments contained the
still, for example, use label 〈110〉 for conditioning and ex- elements as prescribed by the conditioning labels i.e. com-
pect to generate a level that blends SMB and KI. The model pared the label for a generated segment with the condition-
architecture was similar to that used for game elements. We ing label used to generate it. The label for the generated
duplicated the number of training segments for KI to better segment was determined using the same method for assign-
match the number of segments for SMB and MM. ing labels to training segments. We computed the percentage
In each of the above cases, we trained 3 versions of each of segments for which the output label was an exact match
model, consisting of latent spaces of size 32, 64 and 128. as the one used for conditioning, as well as the percentage
where none of the elements that the label indicated should
Results be present were actually present in the generated segment.
We performed a three-part evaluation focusing on each of We also performed this evaluation for the original level seg-
the conditioning cases described above. Note that for KI and ments from each game by forwarding them through the en-
MM figures, we reuse certain sprites from SMB. Paths for coder and decoder using each label. Percentages of exact and
all games are shown using a Mario character sprite. none matches averaged across all labels are given in Table 1.
Figure 3: Results of game element conditioning for both
Figure 4: Results of game element conditioning for both
generated and training levels of KI (top) with frequencies
generated and training levels of MM (top) with frequencies
of each label in the training data (bottom). X-axis values are
of each label in the training data (bottom). X-axis values are
the integer encodings of the equivalent binary label.
the integer encodings of the equivalent binary label. Results
shown for 16 most frequent labels in the training levels.
For all games and for both random samples and training
Design Patterns
levels, the 32-dimensional model leads to the highest per- Evaluating design pattern conditioning was more challeng-
centage of exact matches. The 64-dimensional models do ing. Unlike for game elements, where input segments could
better in producing lowest percentages of no matches but not be labeled automatically by checking for elements within the
too much better than 32. Interestingly, in most cases the 128- segment, labels for design patterns had to be assigned man-
dimensional model exhibits the worst performance for these ually for each segment since we lack automated methods for
measures. Across games, results for KI are most promising identifying design patterns. Consequently, we could not au-
in terms of both the highest percentage of exact matches and tomatically determine if the label for a generated segment
lowest percentage of cases with no matching elements while matched with the label used to generate it. Moreover, due
results for MM were the worst. The average Exact match to the low number of training segments and high number of
percentages seem low primarily because the model produces labels, training a classifier to determine a segment’s design
few exact matches when conditioning with labels not in the pattern was also not feasible. Thus, we restrict our evalu-
training data. Thus we plotted match percentages for both ation in this case to visual inspection with Figure 6 show-
generated and training segments obtained when using each ing example segments generated by conditioning on differ-
label. Results for the 32-dimensional model for SMB, KI ent labels corresponding to different combinations of design
and MM are given in Figures 2, 3 and 4 respectively. Each patterns. Results are shown for the 8 most common design
plot also depicts the frequency of each label in the training pattern combinations in the training data. While not a robust
data. Labels are shown along the horizontal axis using the evaluation, we see that the labels are reliable in producing
integer representation for the equivalent binary encoding. the indicated design patterns in the generated segment.
For all games, conditioning using labels that appear fre- Blending
quently in the training levels is a lot more reliable. Example To evaluate blending, we randomly sampled 1000 latent vec-
levels generated using a selection of labels for conditioning tors and conditioned the generation of each using each of
using the 32-dimensional models are shown in Figures 5. the 8 possible labels denoting the 8 possible combinations
These show that the affordances of the CVAE enable both of the 3 games. We trained a random forest classifier on the
controllable generation and the generation of novel varia- training segments using the respective games as the class la-
tions of existing content. The top rows in each of these fig- bels for the segment. We obtained a 99.12% classification
ures demonstrate how an existing segment could be edited accuracy using a 80%-20% train-test split. Note there are 2
into a new one by simply changing the conditioning label. label types: conditioning labels appended to latents to con-
This holds promise for co-creative applications as a list of trol generation, and game labels (SMB/KI/MM) predicted
designer preferences regarding the elements they want in by the classifier indicating which game it thinks a generated
generated segments could be converted to binary condition- segment belongs to. For each conditioning label, we then
ing labels and then used to generate the desired segments. computed the percentage of generated segments that were
32-dim CVAE 64-dim CVAE 128-dim CVAE
Label SMB KI MM SMB KI MM SMB KI MM
〈000〉 38.7 18.1 43.2 31 20.3 48.7 41.5 18.2 20.3
〈001〉 3.8 2.4 93.8 2.7 3.7 93.6 3.5 2.9 93.6
〈010〉 0.7 95.5 3.8 1.5 93.6 4.9 0.7 94.5 4.8
Random 〈011〉 6.8 22.9 70.3 7.8 27.5 64.7 10 24 66
〈100〉 97.6 1.4 1 98.8 1.1 0.1 98.9 0.7 0.4
〈101〉 71.9 2.9 25.2 20.7 5.2 74.1 38.1 2.6 59.3
〈000001〉 〈00010〉 〈00100〉 〈01000〉 〈10000〉
〈110〉 86.5 11.8 1.7 59 34.5 6.5 57.4 33.5 9.1
(a) SMB
〈111〉 56.7 10.3 33 32.1 16.8 51.1 45 11.1 43.9
Table 2: For each label, percentage of blended segments
generated using that label, that was classified as the differ-
ent games. Highest percentage classification for each label-
dimensionality pair highlighted in bold.
Random
and only if its corresponding bit in the label is 0. These re-
〈0001〉 〈0010〉 〈0100〉 〈1000〉
(b) KI
sults suggest that segments generated using these labels do
blend the games. Example blended levels are shown in Fig-
ure 7. We see that adding KI or MM labels to SMB segments
makes them more vertical. Similarly, adding SMB labels to
KI and MM gives them a more horizontal progression.
As further evaluation, for each blend label, we generated
Random 1000 segments and computed the E-distance between them
and the original training segments for each game. E-distance
(Székely and Rizzo 2013) measures the similarity between
〈00001〉 〈00010〉 〈00100〉 〈01000〉 〈10000〉 two distributions and has been suggested as a suitable metric
(c) MM for comparing generative models (Summerville 2018). The
lower the E-distance between two distributions, the more
Figure 5: SMB, KI and MM segments generated by condi- similar they are. Thus, for example, we would expect the
tioning the original segment on the left (top) and a random E-distance between original SMB levels and those gener-
vector (bottom) using the corresponding labels, as explained ated using the blend conditioning label 〈100〉 to be the low-
in Figure 1. Changing the label changes the content gener- est among all labels with the value being higher for labels
ated using the same the latent vector. not containing SMB (i.e. 0 in the first label element). For
computing E-distance, we used four tile-based properties:
Density, Nonlinearity, Leniency and Interestingness as de-
scribed in Snodgrass and Sarkar (2020). Results for all 3
classified as SMB, KI or MM based on the classifier’s pre- blend CVAE models are shown in Figure 8, averaged across
dicted game label. When classifying segments generated via the models. Results for SMB and MM are as expected with
conditioning using one of the blended conditioning labels the lowest E-distance for all models being for labels 〈100〉
(i.e. SMB+KI, KI+MM, SMB+MM and SMB+KI+MM), and 〈001〉. For these, E-distance is higher when that game is
we expect predicted game labels to be more spread across not included in the label than when it is, as expected. Com-
the 3 games versus when classifying segments conditioned paring to SMB, we note the sharp drop as the first label bit
using a single game conditioning label. For example, for flips to 1 indicating inclusion of SMB in the conditioning.
segments generated using the SMB label (〈100〉), we ex- Similarly, comparing to MM, we note the E-distance oscil-
pect a very high percentage to be classified as SMB and a late lower to higher as the rightmost bit switches between 1
very low percentage classified as others. For those gener- and 0, indicating inclusion of MM. Interestingly, results for
ated using the blended labels, we would expect predictions KI are not as expected with label 〈010〉 producing the sec-
with more variance. For example, for the SMB+MM label ond highest E-distance and would be worth exploring in the
(〈101〉), we would expect most segments classified as SMB future. Overall however, the E-distance trends suggest that
or MM (but not too many for either) and very few classified conditioning is able to generate levels of different blends.
as KI. Results are given in Table 2 and are in accordance
with expectations. For the original game labels (i.e.〈100〉,
〈010〉 and 〈001〉), a majority of segments (over 93% in all
Conclusion and Future Work
cases) are classified as the corresponding game. This drops We explored conditional VAEs and how their affordances
and is more spread out as the blended labels (i.e. those with can inform level design applications by allowing design-
multiple 1s) are used. Also as expected, predictions are most ers to use labels to control level generation and blending
evenly spread out for 〈000〉 and 〈111〉. Interestingly, in all as well as edit existing levels using different labels. There
cases other than 〈000〉, a game is predicted 10% or less if are several future avenues to consider. While CVAEs en-
Original 〈SU〉 〈MP〉 〈PV-NV-MP〉 〈G〉 〈G-MP〉 〈EH〉 〈EH-MP〉 〈EH-G〉
Figure 6: Example segments generated by conditioning on SMB design patterns. The first segment in each row is from the
game. Every other segment in its row is generated using the same vector as the original but conditioned using the label for that
column. Results shown were generated using the 32-dimensional model. Labels indicate design patterns as defined previously.
SMB
KI
MM
Random
〈000〉 〈001〉 〈010〉 〈011〉 〈100〉 〈101〉 〈110〉 〈111〉
Figure 7: Segments generated by conditioning on blend labels, using an original segment from SMB (first row), KI (second)
and MM (third) and a random segment (last). First, second and third elements of the label correspond to SMB, KI and MM
respectively. Results were generated using the 128-dimensional model. For first 3 rows, bordered segments are originals.
ing labels that are infrequent in the training data. This could
also help blending applications. Currently, the CVAE gener-
ates blended levels using given labels but the actual blended
content is not controllable. Evolution could search for lev-
els that in addition to being conditioned by a specific label,
optimize a desired objective. In this work, we highlighted
different CVAE affordances for controllable PCG but hope
to focus on each affordance more thoroughly in the future.
A user evaluation for design pattern conditioning would be
useful as would playability evaluations. Additionally, this
approach worked with level segments. To generate whole
levels, it could be combined with the approach in Sarkar and
Figure 8: E-distances between original training distributions Cooper (2020a) that learns a sequential segment generation
and distributions of 1000 levels generated using each of the model where generated segments logically follow to create
blend conditioning labels. whole levels. Finally, we intend to incorporate such CVAE
models into game design tools to enable conditional genera-
tion and design, as outlined in Sarkar and Cooper (2020b).
able controllability without having to run evolution, the lat-
ter could still be useful, for example, to generate content us-
References toencoders. In Proceedings of the 11th Workshop on Pro-
Alvarez, A.; Dahlskog, S.; Font, J.; and Togelius, J. 2020. cedural Content Generation in Games.
Interactive constrained map-elites analysis and evaluation of Sarkar, A., and Cooper, S. 2020b. Towards game design via
the expressiveness of the feature dimensions. arXiv preprint creative machine learning (GDCML). In Proceedings of the
arXiv:2003.03377. IEEE Conference on Games.
Bontrager, P.; Roy, A.; Togelius, J.; Memon, N.; and Ross, Sarkar, A.; Summerville, A.; Snodgrass, S.; Bentley, G.; and
A. 2018. Deepmasterprints: Generating masterprints for dic- Osborn, J. 2020. Exploring level blending across platform-
tionary attacks via latent variable evolution. In 2018 IEEE ers via paths and affordances. In 16th Artificial Intelligence
9th International Conference on Biometrics Theory, Appli- and Interactive Digital Entertainment Conference.
cations and Systems (BTAS), 1–9. IEEE. Sarkar, A.; Yang, Z.; and Cooper, S. 2019. Controllable level
Dahlskog, S., and Togelius, J. 2012. Patterns and procedu- blending between games using variational autoencoders. In
ral content generation: revisiting Mario in world 1 level 1. AIIDE Workshop on Experimental AI in Games.
In Proceedings of the 1st Workshop on Design Patterns in Schrum, J.; Gutierrez, J.; Volz, V.; Liu, J.; Lucas, S.; and
Games, 1–8. Risi, S. 2020. Interactive evolution and exploration
Dormans, J. 2010. Adventures in level design: generating within latent level-design space of generative adversarial
missions and spaces for action adventure games. In Pro- networks. In Genetic and Evolutionary Computation Con-
ceedings of the 2010 Workshop on Procedural Content Gen- ference (GECCO).
eration in Games, 1–8. Smith, G.; Treanor, M.; Whitehead, J.; and Mateas, M. 2009.
Fauconnier, G., and Turner, M. 1998. Conceptual integration Rhythm-based level generation for 2D platformers. In Pro-
networks. Cognitive Science 22(2):133–187. ceedings of the 4th International Conference on Foundations
Goodfellow, I.; Abadie-Pouget, J.; Mirza, M.; Xu, B.; of Digital Games, 175–182.
Warde-Farley, D.; Ozair, S.; Courville, A.; and Bengio, Y. Snodgrass, S., and Ontañón, S. 2016. Controllable proce-
2014. Generative adversarial nets. In Advances in Neural dural content generation via constrained multi-dimensional
Information Processing Systems. Markov chain sampling. In 25th International Joint Confer-
Gow, J., and Corneli, J. 2015. Towards generating novel ence on Artificial Intelligence.
games using conceptual blending. In Proceedings of the 11th Snodgrass, S., and Ontañón, S. 2017. Learning to generate
Artificial Intelligence and Interactive Digital Entertainment video game maps using Markov models. IEEE Transactions
Conference. on Computational Intelligence and AI in Games 9(4):410–
Gutierrez, J., and Schrum, J. 2020. Generative adversarial 422.
network rooms in generative graph grammar dungeons for Snodgrass, S., and Sarkar, A. 2020. Multi-domain level
the legend of zelda. arXiv preprint arXiv:2001.05065v1. generation and blending with sketches via example-driven
Guzdial, M., and Riedl, M. 2016a. Game level generation bsp and variational autoencoders. In Proceedings of the 15th
from gameplay videos. In Twelfth Artificial Intelligence and Conference on the Foundations of Digital Games.
Interactive Digital Entertainment Conference. Sohn, K.; Lee, H.; and Yan, X. 2015. Learning struc-
Guzdial, M., and Riedl, M. 2016b. Learning to blend com- tured output representation using deep conditional genera-
puter game levels. In Proceedings of the 7th International tive models. In Proceedings of the 28th International Con-
Conference on Computational Creativity. ference on Neural Information Processing Systems.
Kingma, D., and Welling, M. 2013. Auto-encoding varia- Summerville, A., and Mateas, M. 2016. Super Mario as a
tional Bayes. In The 2nd International Conference on Learn- string: Platformer level generation via LSTMs. Proceedings
ing Representations (ICLR). of 1st International Joint Conference of DiGRA and FDG.
Liapis, A.; Yannakakis, G.; and Togelius, J. 2013. Sentient Summerville, A. J.; Snodgrass, S.; Mateas, M.; and
Sketchbook: Computer-aided game level authoring. In Pro- Ontañón, S. 2016. The VGLC: The video game level cor-
ceedings of the 8th International Conference on the Founda- pus. In 7th Workshop on Procedural Content Generation at
tions of Digital Games. 1st Joint International Conference of DiGRA and FDG.
Mirza, M., and Osindero, S. 2014. Conditional generative Summerville, A.; Snodgrass, S.; Guzdial, M.; Holmgård, C.;
adversarial networks. arXiv preprint arXiv:1411.1784. Hoover, A. K.; Isaksen, A.; Nealen, A.; and Togelius, J.
Paszke, A.; Gross, S.; Chintala, S.; Chanan, G.; Yang, E.; 2018. Procedural Content Generation via Machine Learning
DeVito, Z.; Lin, Z.; Desmaison, A.; Antiga, L.; and Lerer, (PCGML). IEEE Transactions on Games 10(3):257–270.
A. 2017. Automatic differentiation in PyTorch. In NIPS Summerville, A. 2018. Expanding expressive range: Evalu-
Autodiff Workshop. ation methodologies for procedural content generation. In
Sarkar, A., and Cooper, S. 2018. Blending levels from dif- 14th Artificial Intelligence and Interactive Digital Enter-
ferent games using LSTMs. In AIIDE Workshop on Experi- tainment Conference.
mental AI in Games. Székely, G. J., and Rizzo, M. L. 2013. Energy statistics: A
Sarkar, A., and Cooper, S. 2020a. Sequential segment- class of statistics based on distances. Journal of statistical
based level generation and blending using variational au- planning and inference 143(8):1249–1272.
Togelius, J.; Preuss, M.; Beume, N.; Wessing, S.; Hagel-
back, J.; and Yannakakis, G. N. 2010. Multiobjective explo-
ration of the Starcraft map space. In 2010 IEEE Symposium
on Computational Intelligence and Games (CIG).
Torrado, R. R.; Khalifa, A.; Green, M. C.; Justesen, N.;
Risi, S.; and Togelius, J. 2019. Bootstrapping conditional
GANs for video game level generation. arXiv preprint
arXiv:1910.01603.
Valls-Vargas, J.; Zhu, J.; and Ontañón, S. 2017. Graph
grammar-based controllable generation of puzzles for a
learning game about parallel programming. In Proceedings
of the 12th International Conference on the Foundations of
Digital Games, 1–10.
Volz, V.; Schrum, J.; Liu, J.; Lucas, S. M.; Smith, A.; and
Risi, S. 2018. Evolving Mario levels in the latent space
of a deep convolutional generative adversarial network. In
Proceedings of the Genetic and Evolutionary Computation
Conference, 221–228. ACM.
Yan, X.; Yang, J.; Sohn, K.; and Lee, H. 2015. At-
tribute2image: Conditional image generation from visual at-
tributes. arXiv preprint arXiv:1512.00570.