=Paper=
{{Paper
|id=Vol-1907/13_mici_nelson
|storemode=property
|title=Mixed-Initiative Approaches to On-Device Mobile Game Design
|pdfUrl=https://ceur-ws.org/Vol-1907/13_mici_nelson.pdf
|volume=Vol-1907
|authors=Mark J. Nelson,Simon Colton,Edward J. Powley,Swen Gaudl,Peter Ivey,Rob Saunders,Blanca Pérez Ferrer,Michael Cook
|dblpUrl=https://dblp.org/rec/conf/chi/NelsonCPGISFC17
}}
==Mixed-Initiative Approaches to On-Device Mobile Game Design==
https://ceur-ws.org/Vol-1907/13_mici_nelson.pdf
Mixed-Initiative Approaches to
On-Device Mobile Game Design
Mark J. Nelson Abstract
Simon Colton Playing casual games is a wildly popular activity on
Edward J. Powley smartphones. However, designing casual games is done
Swen E. Gaudl by a smaller group of people, usually on desktop com-
Peter Ivey puters, using professional development tools. Our goal
Rob Saunders is to bring these activities closer together, in terms of
Blanca Pérez Ferrer who does them and how they do them. Our Gamika
Michael Cook Technology platform is a 2D physics-based mobile
game design environment. It comprises a 284-
The MetaMakers Institute dimensional parametric design space, and poses mobile
Falmouth University game design as the problem of navigating this space.
Penryn, Cornwall, UK We have built three mobile apps thus far to experiment
metamakersinstitute.com with on-device, mixed-initiative navigation of the
Gamika design space and some of its subspaces. We
describe these apps here in terms of the initiatives that
go into making a game with them, and how these are
split between people and underlying AI software. Our
overall goal is to democratise game design, so that an-
yone and everyone can make casual games directly on
their mobile phones or tablets.
Author Keywords
Mobile games; mixed-initiative interfaces; automated
game design; automated playtesting; design spaces.
ACM Classification Keywords
Copyright © 2017 for this paper is held by the author(s). H.5.m. Information interfaces and presentation: Miscel-
Proceedings of MICI 2017: CHI Workshop on Mixed-Initiative Creative
Interfaces
laneous
Introduction ling the physics engine, player interactions and scor-
At The MetaMakers Institute and its associated spinoff ing/win-conditions. Physics parameters expose common
company MetaMakers Ltd. (metamakersinstitute.com), features of a 2D physics engine: object spawn
we are building apps for on-device, mixed-initiative rates/locations, collision responses, attractive/repulsive
design of games for mobile devices (smartphones and forces, etc. Interaction parameters specify how players
tablets). We explain here our overall approach in build- interact with the physics world, such as when and how
ing the Gamika Technology platform that provides the objects respond to the player tapping or dragging on
basis for our work, and three apps built on that plat- the screen. Scoring and win-condition parameters spec-
form: Cillr, Wevva and No Second Chance. ify how events impact the game outcome (the more
narrowly conceived “rules” of the game). A more de-
Our goal is to allow users to create mobile games di- tailed parameter overview is given in [4, Section III].
rectly on the devices that they play these games on.
Today, many people play casual games, but a much Games for the Gamika platform are encoded in 284-
smaller number of people design such apps, and they parameter chromosomes (the term is borrowed from
tend to do so on traditional computers, using environ- evolutionary algorithms, as automated game genera-
ments such as Unity, XCode or Android Studio that are tion is a goal), augmented with data such as graphical
Figure 1: Four Gamika games entirely dissimilar to the context in which games are and sound assets. Given a chromosome, the Gamika
played (and much less accessible). platform can run the game via an interpreter that al-
lows run-time changes to the game specifications (see
This goal of on-device creation is why we are building Fig. 1 for a few examples).
mixed-initiative, co-creative design tools [5]. On-device
design of mobile games must balance two issues: giv- Cillr: Navigating the full design space
ing users enough control to feel ownership of their out- Given the Gamika Technology platform, the problem of
put, but automating enough aspects of the design exer- on-device navigation of a high-dimensional game-
cise so that making games on a smartphone feels like design space is still far from trivial. But compared to
an enjoyable, empowering exploration of design possi- the open-ended problem of full game design, we be-
bilities, not cumbersome small-screen programming. lieve it serves as a more suitable starting point for the
Hence our goals are more aligned with the genre of affordances of mobile devices.
creative apps dubbed “casual creators” [2] than with
visual-programming tools. Equally importantly, design spaces can be both manual-
ly and automatically navigated, allowing for mixed-
Our approach begins by parameterising 2D physics- initiative design. We are working on both interface- and
based games. The basis of the Gamika Technology plat- automation-oriented solutions to on-device design-
form is a 2D game engine parameterised by 284 fea- space navigation, and experimenting with using the two
Figure 2: Cillr design panels
tures that we have identified as core to a diverse range together. Cillr, our in-house app for building Gamika
of casual games. This set includes parameters control- games (Fig. 2), implements baseline versions of both.
The simplest way of manually navigating a 284- time hunting for which slider to change to make some-
dimensional design space to give the user 284 sliders, thing specific happen. Furthermore, even after having
with which they can set each parameter. While simplis- found the desired parameter, it can be difficult to un-
tic approach, this does work fairly effectively in Cillr. derstand why the game didn’t change as expected.
The 284 sliders are grouped into categories with related
functionalities to make them more discoverable In a preliminary user test with game-design under-
(spawning-related sliders, collision-related, etc.). graduate students, we found them somewhat frustrated
by the experience of using Cillr to make games. Inter-
The simplest way of automatically navigating a large face complexity was one issue, but more importantly,
parameter space is to randomise the parameters. How- the difficulty of understanding the high-dimensional
ever, we have found that this produces too low a yield design space made it hard for these initial testers to
of playable games, and hence Cillr mutates subsets of grasp what they wanted to do in the app, and how they
parameters from existing games instead. Randomly would begin to do it. Therefore, rather than focus ini-
mutating multiple sets to produce a new random game, tially on improving Cillr’s interface, we have instead
and then trying to figure out what it is, can be a fun focused on producing design tools for more cohesive,
interaction loop. If you aren't a researcher interested in lower-dimensional design subspaces, still on top of the
design spaces, however, the proportion of playable overall Gamika Technology platform, but not exposing
games remains too low for the mutation approach in the entire design space at once.
Cillr to be ready for end-user consumption.
Carving out cohesive subspaces
Besides producing Gamika chromosomes (both manual- The next phase of our research has looked at restricting
ly and with randomisation), Cillr includes editing tools the larger Gamika design space to more cohesive sub-
for graphical elements such as sprites, level layout, and spaces, which exposes more comprehensible on-device
lighting, so complete games can be produced, including design spaces by specialising interfaces and automating
games with level progressions and multiple levels of generative aspects to navigate the subspace.
difficulty. We have used the interface to produce clones
of classic games like frogger, asteroids and space in- Despite all games being 2D and physics-based, the
vaders, as well as a variety of novel casual games (a Gamika space is heterogeneous, with very different
narrated set of design sessions is reported in [1]). kinds of games available within its parameters; some
puzzle-like, others meditative, others arcade-style ac-
We do not, of course, claim that an on-device mobile tion, etc. Cohesive subspaces share enough features
game design tool with 284 sliders and parameter ran- such that navigating the design space feels more akin
domisation is the solution to the problem of democra- to designing game levels, or game variants, with more
tising game design. But as an initial baseline, Cillr is understandable relationships between parameter
perfectly usable, at least by experts. Its main drawback changes and changes in gameplay behaviour (though
is that it is complicated to navigate, and requires some often still with complex and emergent aspects).
Once we've identified a design subspace, the research We used Cillr to produce three variations of Let it Snow
question then becomes: given this design subspace, called Rain Rain, Jack Frost and Slush Slosh, each re-
can we understand its space of variation well enough to quiring different tactics and skills. These winter games
build user-interface and generative components that will be paired with games representing additional sea-
match with its salient features, and employ those to sons, for released as an iOS game, Wevva (Fig. 3). This
build an enjoyable, mixed-initiative app for designing app further includes two aspects that are not common
(and playing) games or levels in that subspace? in casual games: (a) an AI player for each game that
can assist novice players, and (b) a design screen ena-
Below, we describe the first two subspaces we’ve inves- bling players to generate levels in a semi-random way,
tigated, and the corresponding mobile game-design and tweak them to get balanced variations. The AI
apps, namely Wevva and No Second Chance. player appears on-screen as a gloved hand that taps
the blue balls to keep clusters of four from forming
Wevva (Fig. 3, bottom right), implementing one part of a win-
Using Cillr, we made a relatively addictive four-in-a-row ning strategy. A slider lets the player change the level
game called Let It Snow, where snow and rain pour of AI assistance. At 50%, it feels like having an in-
down from the top of the screen (as white and blue game partner helping out. At 100%, the game is quite
Figure 3: Wevva rules (top) and balls respectively). When four or more white balls clus- different, as the AI player takes care of one aspect of
gameplay (bottom) ter together, they explode and the player gains a point the game (avoiding losing points), freeing the player to
for each in the cluster. Each white ball that explodes is concentrate purely on gaining points.
replaced by a new one spawned at the top, with a max-
imum of 20 on screen at any one time. Likewise with The design screen (Fig. 4) exposes the following ele-
blue balls, except the player loses points for them. ments of the game design to the player: (a) the sizes
Players can interact with the game by tapping blue balls at which clusters of balls explode (b) the scores at-
to explode them, losing one point in doing so. tached to clusters exploding and the player tapping (c)
the size of the balls (d) the maximum number of balls
While the game rules are straightforward, we have of each type allowed (e) the design of the grid, (f)
found it to be difficult and require puzzle-solving strat- physical properties of the environment, namely bounci-
egies as well as quick reactions. There is a grid struc- ness and noise, (g) spawning regions for both types of
ture which collates the balls into bins, and the best way balls, and (h) what happens when the player taps the
to play the game involves trapping the blue balls in balls – both actions and scoring consequences.
groups of twos and threes at the bottom, while the
whites are exposed and are continually refreshed There is a random generation button which will set the-
through cluster explosions. Occasionally, when all blues se parameters in a varied way, but designed so that the
are trapped in small clusters, only whites will spawn, clustering explodes are balanced in terms of expected
which is akin to snowing (hence the game’s name) and score. We achieved this by running online simulations
Figure 4: Wevva design panels
is a particularly pleasing moment to aim for. of novice players and recording the number of times
that clusters of each size and type occurred. Initial ex- style and a variety of physics parameters to be
periments with the design screen have indicated that changed. Since what is fixed about No Second Chance
the space exposed by the above parameters, while games is the control and scoring mechanism, new
vast, does not contain hugely varied game types. How- games are made by varying physics, spawning and
ever, we have used it to make games which differ sub- scoring options, which can produce very different game
stantially from the four preset games, e.g., involving dynamics and mechanics.
juggling balls, or trapping and tapping them, etc.
To demonstrate the types of games that can be pro-
No Second Chance duced (and to provide an initial challenge), the app
Again using Cillr, we designed a game of patience and comes with 100 games we designed using this inter-
concentration, Pendulands. Here, balls move in a pen- face, which we’ve categorised into three primary types
dulum-type motion and annihilate each other if they of challenges: skill games, where the primary challenge
collide; the player must catch five of them by hovering is dexterity, ingenuity games, where the primary chal-
under them with a large round target until they stick. lenge is figuring out a specific trick or strategy, and
patience games, which involve waiting for the right sit-
By varying parameters within this theme, we discov- uation to arise and capitalising on it accordingly.
Figure 5: No Second Chance ered that a whole set of Pendulands variants (or levels)
can be created. The fixed elements defining this sub- The generation button creates a new game via an evo-
space of Gamika games are: the player always controls lutionary process. In particular, pairs of existing games’
the target by dragging, and must catch five balls on the chromosomes are randomly crossed over, then filtered
target. Within these parameters, very different types of using static heuristics to reject clearly bad candidates.
challenges can be created (see Fig. 5 for examples). The first four candidates that pass the filter are auto-
playtested on the device in a split-screen view (Fig. 6,
No Second Chance is our third app, built around this bottom) that plays them at 8x speed for 5 seconds, the
space of games. The name comes from a meta-game equivalent of 40 seconds of game time. We want
mechanic: players can send games to each other in games to be playable but not too easy, so the app
such a way that they are deleted if the receiver doesn’t chooses the game that the playtester was able to catch
beat the game on first playing (in five minutes). This the most balls on, without being able to catch all five.
emphasises the “disposable” nature of games in a gen- (This split-screen visualisation of playtesting isn’t strict-
erative space, where part of the challenge is exploring ly necessary, but we are exploring the entertainment
the space of games and figuring out how each one value of “Hollywood AI” that visually externalizes to
works when first encountering it. users what the apps’ AI components are doing.)
As with Wevva, a design screen (Fig. 6, top) lets play- We have been conducting playtests with the beta app
Figure 6: Design interface (top)
ers make new No Second Chance games. It is laid out to improve both the design interface and generator. A
and auto-playtester (bottom)
as a hierarchical menu, with submenus allowing visual series of playtests in a local school (Camborne Science
and International Academy) have been particularly Second Chance), as AI assistant players (in Wevva), as
helpful, as the students have proven adept at master- fixers for broken player-designed games [4], and even
ing the app and providing useful suggestions. Taking a as performers in standalone art installations in which
more explicitly learning-technology turn, since game the autoplayer both designs and plays new games (as
design in No Second Chance is essentially modification per the installation called I Create, You Destroy shown
of physics parameters to produce new types of game- at the first Arts as Games/Games As Arts festival).
play dynamics, we have also written a series of com- Since the auto-playtesters serve multiple roles, and
panion lessons that introduce game design and basic often need to play in a way that is readable by users,
physics through No Second Chance design exercises. we currently handcraft them out of modular heuristics
in each domain, rather than using off-the-shelf but
Conclusions and future work black-box general game playing algorithms such as
Our goal is to democratise mobile game design by MCTS or deep learning. We are interested, however, in
building on-device design tools, so players can design automatically inferring these kinds of modular, readable
new games in the same setting in which they play heuristics, along the lines of [3].
them. Our view is that doing so requires building tools
for mixed-initiative navigation of design spaces, where Acknowledgements
player/designers have control over their designs but This work is funded by EC FP7 grant 621403 (ERA
also enjoy the benefits of automated and semi- Chair: Games Research Opportunities). We are grateful
automated exploration of these design spaces. for the feedback provided by our alpha/beta testers.
To summarise our design strategy: (a) The Gamika References
Technology platform parameterises game design so 1. Simon Colton, Mark J. Nelson, Rob Saunders, Ed-
that it becomes navigation of design possibilities rather ward J. Powley, Swen Gaudl, Michael Cook. 2016.
Towards a computational reading of emergence in
than programming (b) we carve out coherent design-
experimental game design. In Proc. CCGW 2016.
space subsets in which the relationship of parameter
2. Kate Compton and Michael Mateas. 2015. Casual
changes and game design is more intuitive, and (c) we
creators. In Proc. ICCC 2015, 228-235.
build mixed-initiative apps mapping design interfaces,
automated game playtesters, and game generators 3. Fernando de Mesentier Silva, Aaron Isaksen, Julian
Togelius, Andy Nealen. 2016. Generating heuristics
onto each subspace. Our first two apps built on such
for novice players. In Proc. CIG 2016, 158-165.
subsets, Wevva and No Second Chance, are discussed
4. Edward J. Powley, Simon Colton, Swen Gaudl, Rob
in this paper, as is Cillr, an internal prototype app that
Saunders, Mark J. Nelson. 2016. Semi-automated
targets the entire Gamika design space. level design via auto-playtesting for handheld cas-
ual game creation. In Proc. CIG 2016, 372-379.
Modular, reusable automated playtesters are a compo-
5. Georgios N. Yannakakis, Antonios Liapis, Constan-
nent worth elaborating on. They take initiative in vary- tine Alexopoulos. 2014. Mixed-initiative co-
ing ways: as evaluators during game generation (in No creativity. In Proc. FDG 2014.