, , Alvarez. et al
such as level design [
1
], and to create meaningful interactions with
depth and context [
10
,
31
]. Thus, narrative in games has been a
focus of study [
2
,
18
,
55
] and its generation has been approached
in dierent ways and with dierent techniques [8,9,21,46,49].
Patterns are a common approach to narrative and other facets.
The focus then has been on extracting common narrative aspects
to ease the identication, encoding, and generation of narrative in
dierent forms [
5
,
12
,
16
,
30
,
52
,
53
]. However, it remains a challenge
to dene certain narrative aspects more aligned with the structure
and overarching goals of the game given what type of content is
generated, as well as using these to design and compare among
games. One approach would be to change the abstraction level at
which the narrative is designed. Instead of focusing on the details,
quests, or plot, one could focus on the structure. Narrative structures
can be used to describe how a story is to be developed, as argued by
Barthes [
11
], and to create an abstract representation that reveals
common structures among them, such as Propp’s 31 “narremes" [
41
].
One approach to generate narrative structures is TropeTwist [
5
],
which uses tropes, narrative conventions found across many media
types [19,50], as patterns to design these structures.
This paper presents Story Designer, a mixed-initiative co-creative
(MI-CC) narrative structure tool built on top of the Evolutionary
Dungeon Designer (EDD) using the TropeTwist system. Story De-
signer uses tropes as building blocks for designers to compose
complete narrative structures by interconnecting them in graph
structures called narrative graphs. Story Designer lets designers
create narrative graphs and assist them with a suggestion grid that
uses the Interactive Constrained MAP-Elites (IC MAP-Elites) [
6
]. By
having an MI-CC system to design narrative structures, designers
could ideate and prototype their structures while the system adapts
and suggest novel narratives, making use of patterns, optimizing
coherence, and situating the narrative structures along dimensions
of interest for designers. At the same time, IC MAP-Elites can take
advantage and use the designer’s structure as a proxy to evaluate
subjective characteristics such as interestingness, which has been
the subject of several studies [32,42,44].
As Story Designer is implemented in EDD and based on the link
between level design and narrative, we make use of the designed
dungeon to create constraints over the narrative generation, eec-
tively intertwining both facets. We assess Story Designer with four
controlled and simulated experiments, three premade structures
of dierent games, and one step by step design that showcase the
possibilities within the system. All experiments were tested with
and without level design constraints and using a pair of dimen-
sions and all dimensions during the search. Our results indicate
that IC MAP-Elites have consistency and stability in generating
content and that delimiting the search space with additional level
constraints, while limiting the diversity and generation of complex
structures, guides better the search.
2 RELATED WORK
There is by now a large body of research on procedurally generating
various types of game content [
45
]. While the literature on PCG
in general is far too voluminous to survey here, it should be noted
that PCG methods of dierent kinds have been developed for a
wide variety of content, not just game levels. Narrative, quests,
and plots have been generated using dierent approaches such
as planning [
54
], grammars [
16
,
23
], machine learning [
49
], and
patterns [
5
,
12
,
53
]. Further, several approaches have been proposed
to generate multiple facets of games, in particular level geometry
together with rules, music, lighting, sound etc [
15
,
21
,
24
–
26
,
33
,
35
,
51
]. More relevantly to the current project, several papers have
proposed ways of co-generating narrative and levels [
3
,
10
,
17
,
23
].
In tandem with research on automatically and autonomously
generating game content and narrative, there has been a consid-
erable amount of work “mixed-initiative” systems, which allow a
human designer to co-create content with algorithms. In the do-
main of level generation for games, a number of systems have been
developed that allow a human to receive suggestions, feedback,
or constraints from an AI systems. These include systems for co-
creating platform game levels [
48
], strategy maps [
34
], and certain
aspects of narrative [28,29].
The core algorithm employed in the current paper is MAP-Elites,
a quality-diversity algorithm that seeks to illuminate a space of pos-
sible problem solutions [
37
]. While essentially a type of evolution-
ary algorithms, MAP-Elites, like other quality-diversity algorithms,
do not seek to nd a single best solution but rather a set of solutions
that vary along certain specied measures. The measures dene a
grid, where each cell is the best solution that has been found within
certain values of the measures. These measures can be dened in
many ways; for game levels, they might include the density of a
level, its diculty for a particular type of agent, its symmetry etc.
MAP-Elites has been used in multiple recent AI-based game design
systems [4,7,13,14,27].
3 STORY DESIGNER
Story Designer is a new system integrated in EDD, which presents a
visual interface for mixed-initiative narrative structure generation.
It makes extensive use of the TropeTwist system as foundation to
build narrative graphs and assess them by identifying trope patterns.
The user manually designs a story structure by adding and intercon-
necting nodes in a graph, which seeds an evolutionary algorithm
(EA) that generates story structure suggestions that can be incor-
porated into the user’s design. This continuous co-creative design
process implements the Interactive Constrained MAP-Elites (IC
MAP-Elites) approach presented in [
4
], providing quality-diverse
suggestions across several feature-dimensions.
Story Designer is interconnected with the level design facet in
EDD. This means that the narrative graphs that can be developed
and that can be generated and suggested are constrained by the
content that exists in the levels. For instance, if the designer adds
two NPCs besides the Hero, then the system could at most, use
three character nodes to represent them, or if the designer adds a
boss enemy and a quest item, this would mean that the boss enemy
could be represented as one of the villain nodes (e.g., Enemy, Big
Bad, or Dragon) and the quest item as a possible Plot Device.
3.1 TropeTwist
TropeTwist [
5
] is a system that uses tropes [
19
,
22
,
43
,
50
], narrative
conventions easily recognizable by the audience, as patterns that
combine to compose narrative structures. These structures dene
generic aspects of a story, leading to the identication of events,