=Paper=
{{Paper
|id=Vol-3926/paper7
|storemode=property
|title=MiniStS: A Testbed for Dynamic Rule Exploration
|pdfUrl=https://ceur-ws.org/Vol-3926/paper7.pdf
|volume=Vol-3926
|authors=Bahar Bateni,Jim Whitehead
|dblpUrl=https://dblp.org/rec/conf/exag/BateniW24
}}
==MiniStS: A Testbed for Dynamic Rule Exploration==
https://ceur-ws.org/Vol-3926/paper7.pdf
MiniStS: A Testbed for Dynamic Rule Exploration
Bahar Bateni1 , Jim Whitehead1
1
University of California, Santa Cruz (UCSC), Santa Cruz, CA 95064
Abstract
Specialized testbeds are valuable tools in game AI and procedural content generation (PCG) research, providing controlled environments
to test and evaluate new algorithms and explore game mechanics. This paper introduces MiniStS, a research testbed modeled after the
rogue-like card game Slay the Spire (StS), tailored for studying game-playing AI and procedural generation. MiniStS features a dynamic
rule system where in-game cards can modify game rules, providing a rich context for research. We argue that such an environment offers
distinct research opportunities by revealing insights into game design, enabling the exploration of rule combinations and synergies, and
advancing AI capabilities. We define possible applications for MiniStS in terms of generalized game-playing that requires understanding
and adaptability to the dynamic changes to the rules, and card generation with a focus on a deepened exploration of rule synergies.
Keywords
Collectible Card Games, General Game-playing, Automated Game Design, Procedural Content Generation
1. Introduction Dynamic Ruleset
Designer
Game AI and procedural content generation (PCG) research Rule Rule Rule Rule Rule
frequently benefits from specialized testbed environments
that are tailored to specific research needs. These environ-
ments provide controlled settings where new algorithms can
be developed and evaluated. Furthermore, they can enable
an examination of game mechanics and design principles. Player Playthrough
One prominent example of this is Infinite Mario Bros
Game State
[1], a research testbed designed to emulate the classic plat-
former Super Mario Bros. Infinite Mario Bros has enabled re-
searchers to explore areas such as level generation [2, 3, 4, 5], Figure 1: The structure of a game with a dynamic rule system.
game-playing AI [6, 7], accessibility [8], visualization of In such a game, the designer defines a set of possible rules or
playtesting data [9], and, venturing outside of games, soft- design elements in the game that can be combined together. The
ware repair [10]. GVGAI [11] is another testbed environ- game then allows the player to manipulate the design of the
game through playing the game and activating certain rules.
ment which implements a variety of 2D arcade games, such
as Space Invaders, Pong, Solarfox, and more. Other research
environments include implementations of Angry Birds [12],
Overcooked [13], and Baba is You [14]. 1.1. Dynamic Rule System Definition
In this paper, we introduce MiniStS,1 a simplified Python
We define games with dynamic rule systems as games in
implementation of the rogue-like card game Slay the Spire
which the rules can be changed by playing the game. One
(StS) [15]. MiniStS serves as a dedicated research environ-
of the most interesting examples of this is the game Baba is
ment for investigating various aspects of game AI and pro-
you [16]. In this game, the player can move blocks around a
cedural generation of cards. Similar to other collectible card
grid-based level in a Sokoban style game. In contrast with
games, Slay the Spire includes cards that can affect the de-
Sokoban, here the blocks can also represents game design
sign of the game by enabling or modifying certain rules of
concepts and elements in the level. For example, if the player
the game. Because of this, it demonstrates characteristics
moves the three blocks representing "Flag", "Is" and "Win"
that allow interesting research avenues.
together, it results in activating a rule in the game were the
One of these characteristics is that a game-playing agent
player touching the flag will result in winning the level.
in this game is required to have reasoning capabilities re-
Baba is You is not the only game that utilizes a dynamic
lated to game design, since it needs to manipulate these
rule system. Collectible card games often include cards
playable rules to its own advantage. Further, agents should
which have game design implications. These cards can en-
be generalized enough to play the game while rules are dy-
able or modify the rules of the game to some extent either
namically changing and adapt its strategies accordingly. As
temporarily or for the duration of the battle.2,3 In these
for a PCG agent, it is necessary to not fixate on the context
games, the players are invited to think about the conse-
of a single ruleset, but to consider all the possibilities arising
quences of activating or modifying different rules in the
from combining different rules in different ways. We discuss
game and use these changes strategically to their advan-
these properties and more in Section 1.2.
tage.
Another genre of games that have many examples of us-
11th Experimental Artificial Intelligence in Games Workshop, November ing a dynamic rule system are roguelikes. The definition
19, 2024, Lexington, Kentucky, USA.
$ bbateni@ucsc.edu (B. Bateni); ejw@ucsc.edu (J. Whitehead) 2
An example of such a card in Hearthstone is Cold Feet which causes
https://bahar.bateni.org (B. Bateni); https://users.soe.ucsc.edu/~ejw/ enemy minions to cost 5 more next turn. This card modifies the initial
(J. Whitehead) rule of "Card A costs X" for one turn.
0000-0002-0701-0311 (B. Bateni); 0000-0002-6887-7330 3
An example of such a card in Magic: the Gathering is Elder Mastery
(J. Whitehead) which enchants a creature so that whenever it deals damage to a player,
© 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License
Attribution 4.0 International (CC BY 4.0). that player is forced to discard two cards. This cards adds a new rule
1
The code is available at https://github.com/iambb5445/MiniSTS to the game for a certain creature.
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
of a roguelike game has been a point of debate in the com-
munity.4 However, one of the main goals of this genre is a Design Gameplay
sense of exploration and a gameplay that varies every run.
Toward this goal, there are numerous examples of rogue-
likes that include rule-altering items in the game. The game
uses these alterations to not only provide a strategic aspect Rules
to the game, but also enable various possibilities in the way
the game is played. Examples of this type of gameplay can Rules
be seen in roguelikes such as the Binding of Isaac [18], Hades
[19], Noita [20], Risk of Rain 2 [21] and more.
Figure 2: The design space resulting from a single set of rules
One important note is that the distinction between games
in a dynamic rule system. The designer initially defines a set of
with static rule systems and dynamic rule systems is not composable rules in the game during the game design process.
always pronounced. For example, a first person shooter The player then dynamically changes the set of active rules during
(FPS) can allow the player to choose between weapons such gameplay by making in-game decisions. As a result, the player
as an sniper or an assault riffle, which affects the design to some extent collaborates on game design decisions.
of the game and subsequently the way the game is played.
However, the extent of these changes and the variety of the
possible designs that can be created by these choices are
more limited. Dynamic rule systems specially involve rules
that can be combined together and create a wide variety of A
unique selections. Additionally, games with dynamic rule
systems rely on the changes to the rules as one of the main
components of the game.
Rules Bad Design* Good Design*
1.2. Dynamic Rule System Research
In the previous section, we describe a dynamic rule sys-
tem and the genres that frequently use such systems. In B
this section, we argue that these games can provide unique
opportunities for research on games.
First, games with dynamic rule systems can help us to dis-
cover novel viewpoints in understanding game design. For
Bad Design* Good Design*
example, in Baba is Y’all [14], the authors describe how “the
game provides an interesting study for procedural level gen- Figure 3: Exploring the effect of adding rule A vs rule B in a
eration where game levels and game rules are intertwined.” dynamic design system. The effect of adding each rule should be
Not only does this introduce a new challenge to level design, viewed in terms of how it affects the space of valid combinations
where the game design and its possible outcomes should in the design space as opposed to in a single fixed ruleset. The
good design and bad design visualizations are assumed according
be considered when designing the levels, but it also sug-
to system’s design quality metric.
gests the possibility that this connection could have been
explored deeper in other games, even ones with static rule
systems.
Second, dynamic rule systems can enable the study of rule for the agent to be able to understand this space to some
combinations in a way that cannot be explored in a static extent in order to steer toward designs that are better suited
rule system. As shown in Figure 2, in a dynamic rule system for the agent’s goals. The challenge for game-playing AI is
the set of available rules creates a space of possible designs to reason about what effects its decisions have on the game
in which every valid point is the result of combining rules by since these decisions relate to game design.
activating them in some order (in a card game, via playing Finally, it’s important to note that research on dynamic
cards). The player choices can then result in moving in this rule systems can be beneficial for improving the develop-
space from one point to the other, resulting in a different ment of such games. In other words, such research has ben-
experience every time. eficial impacts not only on game design and game-playing
When taking the role of the designer and adding a new AI, but also on the development of new games with dynamic
rule to the set of available rules, the effect of this new rule design systems. In roguelikes, random generation and PCG
should be examined on all the possible designs from every systems are often used to create a different experience every
possible combination. This effect can move the design points playthrough and foster a sense of exploration in the player.
in the space and result in a distribution of design points with A PCG system designed specifically for dynamic rule sys-
a lower or higher design quality according to the relevant tem can enhance this design concept by ensuring that the
research metrics. Figure 3 visualizes this characteristic. procedural content aligns more closely with the evolving
This study of rule combinations and synergies is not lim- gameplay, or expand the set of available rules by finding
ited to when an AI assumes the role of the designer. Simi- rules that fit its malleable rule system.
larly, when developing a game-playing AI, it is necessary In a similar fashion, the interactions between dynamically
added rules is important in many games, especially card
4
One popular interpretation of roguelike is the Berlin Interpretation, games like MiniStS. When the cards are designed in a way to
[17] which defines a set of characteristics expected from roguelikes. enable various synergies and interactions between the cards,
However, many of these characteristics, such as turn-based, grid-based,
and ASCII art style, do not apply to a significant number of modern
they enable lenticular design [22], a term introduced by
roguelikes. Mark Rosewater, the head designer of Magic the Gathering.
Lenticular design is when understanding a card is more as applying rhythm-based approaches [2], creating level
difficult for an experienced player compared to a beginner. sections focused on certain mechanics [3], or composing
If the cards include many synergy effects when combined predefined level sections [4, 5].
in certain ways, an experienced player would likely have In addition to its applications in level generation, IMB
to consider all of these effects when playing. On the other has also served as a gameplaying environment acting as
hand, a novice player would not understand these effects a benchmark for artificial intelligence research [7]. This
and hence would not get lost in the complexity of the game. resulted in the introduction of the Mario AI Framework
This is a desirable effect for game designers, since it means [6], which has been utilized in AI competitions to evaluate
that the barrier to entry is low allowing novice players to feel and compare various AI strategies. Beyond these applica-
comfortable playing the game with minimal instructions, tions, IMB has also been used in visualizing and analyzing
while experienced players continue to find new layers of play-testing data [9], exploring accessibility methods [8],
complexity to the game. and contributing to fields outside of game studies, such as
runtime software repair [10].
1.3. Slay the Spire Another research game environment is the physics-based
game Angry Birds. Since the game Angry Birds does not
Slay the Spire is a single-player roguelike card game that has have an open source implementation, Ferreira and Toledo
a strong emphasis on strategic play and card synergies. We implemented a clone of the game for the purpose of explor-
chose this game to implement as the testbed environment ing level generation in this game [12]. This environment,
since it has desirable characteristics from both roguelikes named Science Birds, has been used to explore a variety of
and collectible card games. level generation approaches such as procedurally construct-
First, as with many roguelike games, StS is a single-player ing and combining structures [23], utilizing Generative Ad-
game. The game is played against a set of pre-defined en- versarial Networks (GANs) for level generation [24], and
emies that perform different moves with different proba- using genetic algorithms [24]. Furthermore, Science Birds
bilities independent of the player actions. Therefore, the has been selected as the platform for the GPT4PCG [25]
challenges of understanding the opponent skill level in eval- competition in which the use of Large Language Models
uating game-playing capabilities of an agent can be elim- in generating levels via effective prompt-engineering is ex-
inated in favor of focusing on an understanding of rules plored. The AIBirds [26] competition also uses Angry Birds
and their combinations, both during game-play and during as a testbed for both level generation and game-playing, and
design of these rules. Similarly, the strength of different uses Science Birds for its level generation track.
cards or decks can be evaluated without considering how The IMB and Science Birds testbeds are not the only ex-
well they perform against different opposing decks. amples of re-implementing games for research. Hu et al. list
Furthermore, the core loop of the game allows for an in- other examples of environments developed for research pur-
cremental approach to deck-building. At the end of every poses based on video games [27], some of which have been
battle in StS, the player is asked to choose at most one of significantly simplified to better fit the intended purpose.
three possible cards to add to their deck. The increasing These examples include a simplified version of Overcooked
level of difficulty for each battle and the enemies are care- developed to experiment with cooperative abilities of AI
fully designed in a way that appropriately matches bigger agents [13], a minimal implementation of Real-Time Strat-
and more powerful decks as the game progresses. Because egy (RTS) game mechanics used for microRTS game AI
of this, a game-playing agent defined in this environment competition [28], and an implementation of the Legend of
doesn’t need to necessarily be exposed to the challenges that Zelda: a Link to the Past developed specifically as a PCG
arise from the vast space of possibilities for possible decks. research environment [29]. It’s also worth noting that some
Instead, the agent can be tested in terms of combining rules games such as Starcraft II provide an API for interacting
from a single set of cards in a pre-defined deck, and the with the game which can be used to develop game-playing
game offers various examples of well-designed levels with agents that can directly interact with the game via these
different degrees of difficulty suited for this evaluation. APIs [30].
Additionally, similar to most collectible card games, Slay General Video Game AI (GVGAI) [11] is an example of a
the Spire is turn-based. As a result, game-playing in this general purpose environment that provides a wide variety
environment can be done without any considerations of real- of 2D arcade games as a testbed for evaluating AI agents
time play and mimicking human players in terms of reaction in terms of general video game playing capabilities. The
time. The possibility of eliminating these challenges in StS agents in this environment are expected to not be game-
is desirable since it makes the MiniStS environment better specific, which requires them to be able to play games that
suited for our goals. they have not been trained on. Similar to this goal, MiniStS
also expects the agents to be able to adapt to changes in the
rules of the game. However the StS environment requires
2. Related Work the game-playing AI agent itself to actively make changes
Research in games often utilize game environments that im- to the rules through playing the game. This is an important
plement existing games to explore various aspects of game difference that requires the agent to reason about how a
design, level generation, and artificial intelligence. As men- specific rule would benefit or harm its chance of winning,
tioned earlier, one notable example is Infinite Mario Bros and how best the rules can be composed to fit the agent’s
(IMB), a public domain clone of the classic Super Mario Bros. goals. Furthermore, the card game nature of StS allows
IMB incorporates a simple procedural level generation pro- for the possibility of an agent understanding these rules
cess that combines pre-defined segments to create a level. via language-based reasoning. Our previous work [31] has
One of the ways that researchers have used IMB is by ex- shown that such an agent in this environment can show
perimenting with different level generation methods such long-term reasoning capabilities required in some scenarios.
MiniStS is not the only card game environment avail-
able for research purposes. XMage5 and Hearthbreaker6 Possible Actions
are community-made clones of Magic: the Gathering and
Hearthstone respectively. Ling et al. use the dataset of cards
from both of these environments to perform research on Which To which To which
code generation [32]. Metastone7 is a community-made card? agent? card?
implementation of Hearthstone which Santos et al. use to
explore use of the MCTS algorithm in playing Hearthstone Game State Agent Action
[33]. One important difference between MiniStS and the
community-made environments is our focus on not limiting
MiniStS to the set of existing cards. Since our goal is en- Figure 4: The API required for defining a rule-playing agent in
abling research that explores dynamic rule systems and an MiniStS. The agent should be able to respond to three type of
understanding and adaption to the rules during gameplay, questions given the state of the game and the list of possible
we specifically design MiniStS to facilitate these goals. responses.
Chen and Guy introduce Chaos Cards [34] as a card gen-
eration system. They use the Chaos Card environment,
which has similar rules to Hearthstone, to both generate
Though these two applications can be explored indepen-
playable cards and playtest them by using a 1-step look-
dently, they relate to each other in an interesting way. The
ahead agent. The cards are designed via a specific grammar
dynamic rule-playing agent can be used as a means of eval-
that runs through GIGL [35], which was previously made
uation to playtest the game when new cards are introduced
by the same authors to enable creating executable artifacts
to the game. On the other hand, defining new cards can test
from grammatical descriptions. While this environment is
the agents in terms of generalization since the agent can be
well-suited for the purposes of the authors, our focus lies
provided with new rules to play with without being trained
on cards that allow changes to the rules and can possibly
on those specific set of rules. A virtuous cycle.
have effects beyond the Hearthstone-specific grammar. Ad-
Below we describe the required API to tackle each of
ditionally, MiniStS also removes some of the challenges of
these challenges, as well as our existing examples of agents
playing against an opponent and deck-building to enable a
that respond to each target application.
more direct focus on studying the underlying dynamic rule
system, which we discussed in Section 1.3.
Finally, while not a card game, Baba is You has similarities 3.1. Dynamic Rule-Playing AI
with MiniStS in terms of qualities of the environment. As In MiniStS, we define the term game-playing agent (GPA) as
discussed in Section 1, Baba is You is an interesting research an agent that take the role of the player in the game by mak-
subject since, similar to games with as dynamic rule systems ing any decision that is expected from the player. Any GPA
such as StS, it provides an environment that the player in MiniStS is required to implement three functionalities:
interacts with the rules and changes them during gameplay.
Not only does this create interesting challenges in level 1. Choose a card to play: Given the current state, the
design and intertwines level design and game rules, it also agent should be able to decide which one of the
raises interesting questions in terms of game-playing that available cards to play, or decide to end its turn.
relates to the AI agent’s understanding of how to best play 2. Choose the target (agent): Some of the cards can
with the rules. Charity et al. use an implementation of the target one or more agents in the game, for example
game based on the available code for the initial game jam the player should choose the target for “Strike: Deal
version of the game to research their proposed collaborative 6 damage.” Given the state of the game and a list of
mixed-initiative design system [14]. possible options, the GPA should be able to decide
which option(s) are the target of the card.
3. Choose the target (card): Similarly, some cards can
3. Applications target other cards, such as “Concentrate: Discard 3
In this section, we highlight two types of application for cards. Gain 2 energy.” Given the state of the game
which MiniStS can be used. First, we can use MiniStS to ex- and a list of possible options, the GPA should be able
plore dynamic rule-playing agents, i.e. game-playing agents to decide which option(s) are the target of the card.
that can act in a dynamic rule system, deciding how to play The only requirement for playing the game is for a GPA
the game in ways that possibly result in a change of rules. to be able to respond to these three questions. MiniStS also
The agents should be able to continuously adapt to these allows the agents to simulate the future steps of the game,
changes. Furthermore, the agents should have an under- making it possible to define agents that use this prediction
standing of game design and the effect of rules on the game in their decision making, such as an MCTS agent. Note that
to be able to choose how to manipulate the active rules this simulation is not completely accurate, since any non-
toward their goal. deterministic actions (e.g. shuffling the deck, random choice,
Second, we discuss the use of MiniStS as a card generation etc.) are intentionally forced to use a different seed. This is
environment. Since MiniStS is designed in a way to allow because none of the agents should have perfect knowledge
the definition of new cards by combining existing actions about the non-deterministic events of the game.
and effects or defining new ones, this environment can be In our previous research [31], we developed three differ-
used for designing and playtesting new cards. ent agents within MiniStS and compared them in different
scenarios. First, we showed that a random agent can be
defined by simply choosing one of the options randomly
5
https://github.com/magefree/mage/ for each of these questions, as shown in Figure 5. Second,
6
https://github.com/danielyule/hearthbreaker/
7
https://github.com/demilich1/metastone
we utilized the simulation abilities of MiniStS to showcase
over 1,000 simulations). Running these 1,000 simulations on
a personal laptop with an RTX 3050 GPU and on 50 threads
takes less than 10 milliseconds per simulation on average.
3.2. Card Generation
The second goal of MiniStS is to enable researchers to define
card-generation systems. As we discussed previously, when
Figure 5: Pseudocode showing implementation of an example the cards can alter the rules of the game, card generation
agent (random agent). The agent implements functions to decide requires a careful consideration of rule synergies and the
the next action based on the state of the game and to choose a space of the game design. To achieve this goal, we design the
target between agents/cards when needed. The pseudocode is system in a way to allow versatile combination of different
slightly simplified from the actual implementation. actions and effects. These actions and effects can be from
the existing actions, but a major part of our focus is in not
limiting our system to a pre-defined set of actions. The
a look-ahead agent. The look-ahead agent simulates the actions can also be combined with an AND module. If the
game up to a pre-defined depth and chooses an action for action requires a target, MiniStS forces the card definition
the current move that would maximize the estimated score to include the TO module for defining how this target is
in the future. Finally, we demonstrated how MiniStS can be selected. We discuss these modules in more detail and with
used with a language-based agent by converting the game examples in Section 4.
state into a purely text-based format and sending it to a Additionally, some cards can change the rules to have
large language model (LLM) along with the basic rules of specific effects when certain events occur. To make this pos-
the game. The text-based description for each card is au- sible, we define status effects that visualize these changes in
tomatically generated based on its effects but can also be the rules, similar to how the game visualizes these changes
customized to show a predefined text. All three agents are for the player. We also use an open-ended event system
generalized, meaning that they are not limited to playing that allows for defining broadcast events and subscribing to
only a specific set of cards. These three agents are included them to perform actions that result from the rule change. We
in MiniStS’s source code and can be used to run simulations explain MiniStS components, such as actions, their targets,
in different settings. Our primary focus in this paper is to status effects, and event system in more detail in Section 4.
present MiniStS as a testbed, rather than the agents. Our With this setup, creating a new card in the game can be
previous work [31] provides an example of how MiniStS done by following these steps:
can be used to simulate and compare different agents in 1. Use the set of existing actions and define any new
a dynamic rule system, and includes more information on actions that are required. Each action can access
each of these agents. the state of the game and manipulate it. However,
The testbed that MiniStS provides is uniquely useful in the actions should be defined as atomic operations,
testing rule-playing agents. As discussed in Section 1, one meaning that they are small enough that cannot be
of our main goals with MiniStS is to define an environment broken down further into lower level actions.
in which playing the game can change the way the game 2. If there are any new status effect, define these effects
is played. As such, we call the game-playing agents in this and determine their properties, such as if multiple
environment rule-playing agents, since the agent is playing instances of the same effect would stack, if the effect
with the rules. In other words, the performance of such an is for a certain number of turns, and so on.
agent is tied to its understanding of how these rules change 3. If the card requires any new events or callback on
the game, and whether or not it would benefit from these the existing events, define the new modules.
changes. 4. Combine the actions, and determine any required
MiniStS also includes an agent evaluation module that card parameters (such as cost of the card) or action
allows comparing different agents given a scenario. The parameters (such as amount of damage for the action
scenario defines which cards the agents have in their deck, “Deal x damage”).
and what enemies are they encountering. The evaluation
component then simulates multiple instances of the game Example of card definitions and combining the actions is
in parallel and logs the game events and the response of available in Section 4.
agents to these events. Finally, these logs are imported by To show a simple example of card generation, we im-
another module to calculate a set of metrics such as win-rate, plement a simple grammar-based card generation module.
the distribution of player’s health at the end of the game, The card generation module randomly chooses actions and
whether a certain card or combinations of cards were used possibly combines them by using the AND module. The pa-
by the agent, how many turns the agent took to win/lose rameters for each action are chosen from a set of reasonable
the game, and so on. values that are determined based on examples of existing
The execution time of the simulation is highly dependent cards. The card generation process lacks any evaluation of
on the number of turns it takes for the agent to end the the quality of cards and is also limited to the set of available
game and the agent’s performance. To provide an estimate actions and their combinations, and does not introduce new
of MiniStS’s performance and factor out the time it takes rules into the game. We use the available card generation
for the agent to make decisions, we can consider an agent process not as a strong method for generating cards, but as
that always chooses the first available action. In this case, both an example of how cards can be generated from these
the first battle of the game, using the starting set of cards, actions and to enable a simple way of providing new cards
averages around 5 turns and 23 agent actions (calculated for the purpose of testing rule-playing agents when playing
with never-seen-before cards.
Atomic Actions
Atomic Actions Card Targets Agent Targets
Card Targets Agent Targets
Decorate Decorate Deal Gain Discard
Target
Target
Deal Gain Self Discard Pile Self
Target
Target
Damage Block Self Pile Self
Damage Block Draw
Choose Draw Enemy
Choose Enemy
And Discard Draw Pile
Target
And Discard Draw Choose Pile Choose
Target
Card Card
Card Card All Hand All All
To Apply All Hand All All
Target
… Exhaust
To Apply Status … RandomExhaust Pile Random
Target
Status Random Pile Random
Figure 7: Possible targets for atomic actions. The targets can
Figure 6: Atomic actions. Some actions have a specific type of be either a card or an agent (i.e. the player or enemies). Each of
target (agent or card) and can be decorated with TO. Any action these targets is defined on a set of possible options, and a way of
or combination of actions can be decorated with AND. choosing the option.
4. System Structure • Self: Either the player or the card that includes this
action, based on the type of target (agent or card).
In this section, we describe different components of the Min-
• Choose [from a set of options]: The target is identi-
iStS design with the goal of creating a research environment
fied based on player input.
with a dynamic rule system.
One of our main goals with MiniStS is to enable and • All [from a set of options]: The action with this
streamline the process of defining a new card. To achieve type of target is applied to all the the possible target
this, we analyze a set of existing cards, both from Slay the options.
Spire and from a set of third party cards designed as mods • Random [from a set of options]: The target is chosen
by the StS community. as one or more options from the set of available
options.
4.1. Actions For each of these cases, the set of available options can be
defined in a few different ways. By examining the existing
The first component that we identify in the cards are ac-
cards, we identify the possible set of options for card targets
tions. Actions define what effect the card has on the state
to be either from player’s hand, discard pile, draw pile or
of the game. For example, a “Gain Mana” action means that
exhaust pile. As for the agent, if the target is not the player
when the card is played, it changes the amount of available
(self), it’s from the enemies. We also noticed in a set of third
mana for the player. Some of these actions are visualized in
party cards from community mod packs that the target can
Figure 6.
also refer to all the agents (i.e. both the player and the
The actions are defined as atomic, meaning that when
enemies).
we analyze a card, we break down its effect into indivisible
actions that cannot be further decomposed. We then allow
these atomic actions to be combined together in any order 4.3. Status Effects and Events
by defining an AND decorator that creates an action from Slay the Spire defines a set of status effects, or buffs and
the combination of any two actions. debuffs, which can be used to indicate when an agent is
Finally, many actions in the game require one or more weak (deals less damage for a number of turns), vulnerable
targets. For example, the “Deal damage” action affects one (takes more damage for a number of turns), etc. While we
or more enemies which are selected in a specific way. We can implement only these effects by using specialized code
make this possible by defining actions that need a target. For that only covers these cases, we want the definition of a
each of these actions, the designer should specify how the status effect to be more open-ended and include a flexible
required targets are selected, whether based on player input, infrastructure, since status effects are the foundation of rule-
random selection, or chosen as “all the possible targets.” based cards.
These actions are considered incomplete unless they are So far, we have defined atomic actions and targets which
decorated with TO and a target definition. We discuss the enable many of the available cards in the game to be ex-
target definitions in the next subsection. pressed in MiniStS. However, many of the most interesting
Note that the AND decorator can still be applied before or cards in the game, especially the cards that change the rules
after defining a target. Figure 9 shows pseudocode for two of the game, cannot be defined with the actions alone. For
example cards, and Figure 10 shows how the second card can example, the card “After Image” is defined as “Whenever
be redefine to apply its underlying actions to two, possibly you play a card, gain 1 block.” Instead of having an instant
different, targets. Figure 8 visualizes the two example cards. effect on the state of the game, these cards can have a last-
ing effect that can be triggered at certain points, such as
4.2. Targets whenever a card is played, at the start of every turn, until
the end of this turn, etc.
As mentioned previously, the effect of a card can be ex-
To understand how best we can cover these types of ef-
pressed by a set of atomic actions, some of which require
fects in MiniStS, we first look at how StS communicates
a target. By analyzing the existing cards, we see that StS
these cards and their effects with the player. StS uses status
cards can target agents (i.e. the player or the enemies) or
effects to visualize these changes to the rules of the game.
cards. Each of these targets can be expressed as one of the
For example, the card “After Image”, which we discussed
following types:
above, applies a specialized status called “After Image” on
Gain Discard
Target
Target
To Self And To Choose Hand
Block Card
Deal Apply
Target
Target
And To Choose Enemy
Damage Status
Figure 8: visualization of implementing "Survivor: Gain 8 Block. Discard 1 card." and "Bash: Deal 8 Damage. Apply 2 Vulnerable." Both
cards can be defined by combining atomic actions and targets with AND and TO. However, Bash uses the same target for both actions,
which requires using AND before defining the targets with TO, whereas Survivor have different targets and uses AND after seperately
defining the targets.
Battle
End Turn Play Card
…
Event Event
Before After
Status
Effect
Callback Subscribe
Gain x Block Game State
Figure 9: Pseudocode showing implementation of two example
cards, "Survivor: Gain 8 Block. Discard 1 card." and "Bash: Deal 8 Figure 11: Definition of card “After Image” in MiniStS. The
Damage. Apply 2 Vulnerable." The pseudocode is simplified from applies a special status effect, called “After Image”, to the player.
the actual implementation. The unique callback is subscribed to the event of “Play Card”
to trigger after every time it happens. The callback then, given
the whole state of the game, applies some amount of block to
the player depending on the strength of the status effect which
directly maps to how many times the card was played.
to register the callback to that event. The callbacks can be
subscribed to trigger before or after an event. Figure 11
demonstrates how status effects and the event system can
be used to create “After Image.”
Figure 10: Pseudocode showing implementation of an imaginary The events can also be defined in a way to manipulate
card, "Bash*", which deals 8 damage to one target, and then
a value. For example, the status effect “vulnerable” makes
applies 2 Vulnerable to another, possibly different, target. The
it so that the affected creature takes 50% more attack dam-
pseudocode is simplified from the actual implementation.
age for some number of turns. The vulnerable status ef-
fect is implemented by subscribing to the “attack damage”
event. Basically, whenever the “deal attack damage” action
the player. Each effect also has a set of properties. For exam- is performed in the game, the event triggers and expects
ple, “After Image” status effect does not decrease during the the callbacks subscribed to it to manipulate the amount of
battle, and can stack (i.e. if “After Image” is played twice, damage. Each event system can have any type of input that
the value of status effect is increased to two). it desired, or not have an input at all. Each callback is then
Finally, the status effect performs some behavior when required to specifically get that type of input and returns
the appropriate event triggers. To implement this, we use an the manipulated value.
event system. The event system provides a formal definition Finally, multiple callbacks can be defined on each event.
of how the events and callbacks related to status effects For example, “weak” and “vulnerable” effects are both de-
should be added to the game. Each event has broadcast and fined to trigger when attack damage is happening in the
subscribe functions. Broadcast is used by the entity that game and manipulate its value. However, the order in which
relates to the condition of the event, while subscribe is used
these effects apply completely depend on how the card de- be implemented as combination of two actions by keeping
signer have envisioned the effects. Because of this, it’s im- track of the cost in the game state.
portant for any new callback to be subscribed at the correct StS includes multiple characters, some of which have spe-
order compared to other callbacks. cific mechanics. The current implementation of the game,
which is a simplified version, does not implement these
4.4. Gameplay Loop mechanics. Furthermore, the current version of MiniStS im-
plements some but not all the existing actions, status effects,
The gameplay loop simulates the events of a battle. The and enemies in the game, and only simulates one battle to
event and values of this component are a simulation of the compare agents instead of multiple battles. However, such
events in Slay the Spire. Each turn consists of the following components can be added to MiniStS without any major
events: changes to the code.
1. Gain Mana: The player gains full mana at the start
of the turn. 6. Conclusion
2. Draw a Hand: The player draws cards equal to the
number of pre-defined draw count. If at any times Games with dynamic rule systems are games in which a set
the draw pile is empty, the discard pile is shuffled of composable rules exist which can be activated through
back into the draw pile. playing the game. While the designer is the one who cre-
3. Play (Player Side): The player side is played, mean- ates these rules, the player also participates in game design
ing that the player agent would take its turn, and decisions by playing the game. This unique property, a
then the enemy side would lose any remaining block. key characteristic of collectible card game and roguelike
Note that some damage types would apply after this genres, offers unique research opportunities. AI agents can
event, bypassing enemies’ blocks. be cast into the role of the designer or the player in the
4. Play (Enemy Side): The enemy side is played, mean- system. As a designer, these agents are required to reason
ing that the enemies would take their turn one by about the synergy effect of rules and the distribution of the
one, and then the player side would lose any remain- possible designs that arise from different rules activated in
ing block. each playthrough. As a player, these agents are required to
5. Discard Hand: The cards in player’s hand are then reason about game design in order to combine the rules in
discarded and moved to the discard pile. a way that gets them closer to their goal.
In this paper, we introduced MiniStS, a simplified imple-
The game will continue repeating the events described mentation of Slay the Spire. MiniStS enables researchers to
above every turn until the battle has ended, i.e. the player explore the StS environment and its dynamic rule system
or all the enemies are dead. Note that these events can be in two ways. First, with the role of a player in the system
affected by the cards, e.g. a card can change the number of and as a rule-playing agent, and second, with the role of
draw count, make the cards not get discarded at the end of a designer and as a card generation task. We describe the
the turn, etc. steps with which both of these tasks can be approached in
The gameplay loop is also connected to the game state MiniStS. Finally, we described different components of Mini-
such as the available deck of cards, properties such as draw StS and discussed how each of these components facilitate
count or maximum mana, and state of the battle including a these two applications by making it possible to define new
the list of cards in player’s hand, discard pile, draw pile and cards and rule-playing agent without any major changes to
exhaust pile, and the player and enemies and their respective the code and by using a streamlined process.
states such as HP, mana, block, and so on.
References
5. Limitations
[1] M. Persson, Infinite mario bros, Online Game (2008).
One limitation of the system is that it respects the basic [2] G. Smith, M. Treanor, J. Whitehead, M. Mateas,
rules of the game Slay the Spire, so any changes that are Rhythm-based level generation for 2d platformers,
related to the foundation of the game would require chang- in: Proceedings of the 4th International Conference
ing the system. For example, the order of events (e.g. the on Foundations of Digital Games, FDG ’09, Associa-
player plays first, then the enemies take their actions), the tion for Computing Machinery, New York, NY, USA,
existence of only a single type of mana, or the fact that this 2009, p. 175–182. URL: https://doi.org/10.1145/1536513.
is a single player card game cannot be modified without 1536548. doi:10.1145/1536513.1536548.
applying significant changes to the code. [3] A. Khalifa, M. C. Green, G. Barros, J. Togelius, In-
Another limitation of the system is regarding the types tentional computational level design, in: Proceed-
of input the system expects from the system. As discussed ings of the Genetic and Evolutionary Computation
in Section 4.2, the game expects each action to be applied Conference, GECCO ’19, Association for Computing
to at most one agent or card. Because of this, any action Machinery, New York, NY, USA, 2019, p. 796–803.
that requires a new type of input does not exist in the game. URL: https://doi.org/10.1145/3321707.3321849. doi:10.
For example, one can imagine a card defined as “Remove 1145/3321707.3321849.
one of the status effects of your choosing” which would [4] M. Cerny Green, L. Mugrai, A. Khalifa, J. Togelius,
require a new form of input from the player. Furthermore, Mario level generation from mechanics using scene
actions that are defined on multiple target types do not exist stitching, in: 2020 IEEE Conference on Games
in StS or MiniStS, such as “Discard a card and deal damage (CoG), 2020, pp. 49–56. doi:10.1109/CoG47356.
equal to 5 times its cost.” It’s worth noting that this still can 2020.9231692.
[5] P. Mawhorter, M. Mateas, Procedural level genera- https://magic.wizards.com/en/news/making-magic/
tion using occupancy-regulated extension, in: Pro- lenticular-design-2014-12-15.
ceedings of the 2010 IEEE Conference on Compu- [23] M. Stephenson, J. Renz, Procedural generation of
tational Intelligence and Games, 2010, pp. 351–358. levels for angry birds style physics games, Proceed-
doi:10.1109/ITW.2010.5593333. ings of the AAAI Conference on Artificial Intelligence
[6] S. Karakovskiy, J. Togelius, The mario ai benchmark and Interactive Digital Entertainment 12 (2021) 225–
and competitions, IEEE Transactions on Computa- 231. URL: https://ojs.aaai.org/index.php/AIIDE/article/
tional Intelligence and AI in Games 4 (2012) 55–67. view/12871. doi:10.1609/aiide.v12i1.12871.
doi:10.1109/TCIAIG.2012.2188528. [24] R. Rodriguez Torrado, A. Khalifa, M. Cerny Green,
[7] J. Togelius, S. Karakovskiy, J. Koutnik, J. Schmidhuber, N. Justesen, S. Risi, J. Togelius, Bootstrapping condi-
Super mario evolution, in: 2009 IEEE Symposium tional gans for video game level generation, in: 2020
on Computational Intelligence and Games, 2009, pp. IEEE Conference on Games (CoG), 2020, pp. 41–48.
156–161. doi:10.1109/CIG.2009.5286481. doi:10.1109/CoG47356.2020.9231576.
[8] J. Muñoz, G. N. Yannakakis, F. Mulvey, D. W. Hansen, [25] P. Taveekitworachai, F. Abdullah, M. F. Dewantoro,
G. Gutierrez, A. Sanchis, Towards gaze-controlled R. Thawonmas, J. Togelius, J. Renz, Chatgpt4pcg com-
platform games, in: 2011 IEEE Conference on Com- petition: Character-like level generation for science
putational Intelligence and Games (CIG’11), 2011, pp. birds, in: 2023 IEEE Conference on Games (CoG), 2023,
47–54. doi:10.1109/CIG.2011.6031988. pp. 1–8. doi:10.1109/CoG57401.2023.10333206.
[9] G. Wallner, N. Halabi, P. Mirza-Babaei, Aggregated [26] M. Stephenson, J. Renz, X. Ge, L. Ferreira, J. Togelius,
visualization of playtesting data, in: Proceedings of P. Zhang, The 2017 aibirds level generation competi-
the 2019 CHI Conference on Human Factors in Com- tion, IEEE Transactions on Games 11 (2019) 275–284.
puting Systems, CHI ’19, Association for Comput- doi:10.1109/TG.2018.2854896.
ing Machinery, New York, NY, USA, 2019, p. 1–12. [27] C. Hu, Y. Zhao, Z. Wang, H. Du, J. Liu, Games for
URL: https://doi.org/10.1145/3290605.3300593. doi:10. artificial intelligence research: A review and per-
1145/3290605.3300593. spectives, 2024. URL: https://arxiv.org/abs/2304.13269.
[10] C. Lewis, J. Whitehead, Runtime repair of software arXiv:2304.13269.
faults using event-driven monitoring, in: Proceedings [28] S. Ontañón, N. A. Barriga, C. R. Silva, R. O. Moraes,
of the 32nd ACM/IEEE International Conference on L. H. S. Lelis, The first microrts artificial intel-
Software Engineering - Volume 2, ICSE ’10, Associa- ligence competition, AI Magazine 39 (2018) 75–
tion for Computing Machinery, New York, NY, USA, 83. URL: https://ojs.aaai.org/aimagazine/index.php/
2010, p. 275–280. URL: https://doi.org/10.1145/1810295. aimagazine/article/view/2777. doi:10.1609/aimag.
1810352. doi:10.1145/1810295.1810352. v39i1.2777.
[11] D. Perez-Liebana, S. Samothrakis, J. Togelius, T. Schaul, [29] N. Heijne, S. Bakkes, Procedural zelda: a pcg en-
S. Lucas, General video game ai: Competi- vironment for player experience research, in: Pro-
tion, challenges and opportunities, Proceedings ceedings of the 12th International Conference on
of the AAAI Conference on Artificial Intelligence the Foundations of Digital Games, FDG ’17, Associa-
30 (2016). URL: https://ojs.aaai.org/index.php/AAAI/ tion for Computing Machinery, New York, NY, USA,
article/view/9869. doi:10.1609/aaai.v30i1.9869. 2017. URL: https://doi.org/10.1145/3102071.3102091.
[12] L. Ferreira, C. Toledo, A search-based approach for doi:10.1145/3102071.3102091.
generating angry birds levels, in: 2014 IEEE Confer- [30] O. Vinyals, T. Ewalds, S. Bartunov, P. Georgiev, A. S.
ence on Computational Intelligence and Games, 2014, Vezhnevets, M. Yeo, A. Makhzani, H. Küttler, J. Aga-
pp. 1–8. doi:10.1109/CIG.2014.6932912. piou, J. Schrittwieser, J. Quan, S. Gaffney, S. Petersen,
[13] M. Carroll, R. Shah, M. K. Ho, T. Griffiths, S. Se- K. Simonyan, T. Schaul, H. van Hasselt, D. Silver, T. Lil-
shia, P. Abbeel, A. Dragan, On the utility of licrap, K. Calderone, P. Keet, A. Brunasso, D. Lawrence,
learning about humans for human-ai coordination, A. Ekermo, J. Repp, R. Tsing, Starcraft ii: A new chal-
in: H. Wallach, H. Larochelle, A. Beygelzimer, lenge for reinforcement learning, 2017. URL: https:
F. d'Alché-Buc, E. Fox, R. Garnett (Eds.), Advances //arxiv.org/abs/1708.04782. arXiv:1708.04782.
in Neural Information Processing Systems, vol- [31] B. Bateni, J. Whitehead, Language-driven play: Large
ume 32, Curran Associates, Inc., 2019. URL: https: language models as game-playing agents in slay the
//proceedings.neurips.cc/paper_files/paper/2019/file/ spire, in: Proceedings of the 19th International Con-
f5b1b89d98b7286673128a5fb112cb9a-Paper.pdf. ference on the Foundations of Digital Games, FDG
[14] M. Charity, A. Khalifa, J. Togelius, Baba is y’all: ’24, Association for Computing Machinery, New York,
Collaborative mixed-initiative level design, in: 2020 NY, USA, 2024. URL: https://doi.org/10.1145/3649921.
IEEE Conference on Games (CoG), 2020, pp. 542–549. 3650013. doi:10.1145/3649921.3650013.
doi:10.1109/CoG47356.2020.9231807. [32] W. Ling, E. Grefenstette, K. M. Hermann, T. Kočiský,
[15] Mega Crit Games, Slay the spire, 2019. A. Senior, F. Wang, P. Blunsom, Latent predictor net-
[16] A. Teikari, Baba is you, 2019. works for code generation, 2016. URL: https://arxiv.
[17] Berlin Interpretation - RogueBasin, 2013. URL: org/abs/1603.06744. arXiv:1603.06744.
https://www.roguebasin.com/index.php/Berlin_ [33] A. Santos, P. A. Santos, F. S. Melo, Monte carlo tree
Interpretation. search experiments in hearthstone, in: 2017 IEEE
[18] E. McMillen, F. Himsl, The Binding of Isaac, 2011. Conference on Computational Intelligence and Games
[19] Supergiant Games, Hades, 2018. (CIG), 2017, pp. 272–279. doi:10.1109/CIG.2017.
[20] Nolla Games, Noita, 2019. 8080446.
[21] Hopoo Games, Risk of Rain 2, 2019. [34] T. Chen, S. Guy, Chaos cards: Creating novel dig-
[22] M. Rosewater, Lenticular design, 2014. URL: ital card games through grammatical content gen-
eration and meta-based card evaluation, Proceed-
ings of the AAAI Conference on Artificial Intelligence
and Interactive Digital Entertainment 16 (2020) 196–
202. URL: https://ojs.aaai.org/index.php/AIIDE/article/
view/7430. doi:10.1609/aiide.v16i1.7430.
[35] T. Chen, S. Guy, Gigl: A domain specific language for
procedural content generation with grammatical rep-
resentations, Proceedings of the AAAI Conference on
Artificial Intelligence and Interactive Digital Entertain-
ment 14 (2018) 9–16. URL: https://ojs.aaai.org/index.
php/AIIDE/article/view/13025. doi:10.1609/aiide.
v14i1.13025.