Dread (Barmore et al. 2005) is an award-winning tabletop horror RPG that emphasizes storytelling. One of its innovative mechanics is the use of a questionnaire for character design that helps players develop a personality and history for their characters. Questionnaires vary for different scenarios and characters, but always include “intrusive” questions such as “why are you married if you aren't in love?” or “what's the worst thing you've ever done to a loved one?” In this paper, we discuss work in progress on AutoDread, an implementation of Dread-style questionnaires for video games. The system presents human-authored questions and candidate answers to players to choose between, collecting the implications of those answers in a character model. As implications are accumulated, the system uses a SAT solver to filter questions and answers that are inconsistent with character facts established by previous questions.
recent tradition of “storygames” or “freeform” games, such
as Fiasco
a Jenga tower in lieu of dice to determine outcomes of
character actions. This provides a particularly visceral
implementation of LeBlanc’s aphorism that drama = uncertainty
+ inevitability
Dread’s other novel mechanic is the use of questionnaires for character design. Scenarios are pre-authored by a host (gamemaster) to have a designated set of character roles, such as ship’s captain or camp counselor. Each character role has a pre-authored questionnaire that is answered, in character, by that character’s player before the start of the game. The questionnaire helps the player design their character, establishing useful bits of backstory that can be used as narrative hooks by both player and host to produce drama and tension during gameplay.
Questionnaires cover a range of topics: the character’s motivations, capabilities and limitations, relation to the overall plot, and relationships with other characters. Questionnaires always include so-called “intrusive” questions: questions that get at issues characters would be reluctant to disclose in real life, such as “who else knows you’re a fraud?”
In this paper, we discuss work in progress on AutoDread, an electronic version of Dread-style questionnaires, suitable for use in video games. The system works from a stock of human-authored questions and potential answers, each tagged with their logical implications. For example, a question such as “How long has it been since you saw your brother?” would be tagged with “has brother”. The answer “Not since mom’s funeral” would be tagged “mother dead”, while the answer “We watch the game every Monday night” might be tagged “brother alive” and “loves brother”, or at least “not hates brother”.
toDread: it can’t accept freeform answers from players, requiring them instead to select from a specific list of humanauthored answers to the question. This allows the questionnaire to be machine-readable, at the cost of increased work for the author and decreased player freedom.
As the system asks the player questions, it accumulates
these implications, as well as other facts that follow from the
implications, to gradually create a model of the character
and their backstory. The system uses a SAT solver
AutoDread is an exploratory first-step toward making a generative interactive fiction system mimicking at least some of the player experience of Dread. To be useful for an electronic game, the game would have to have sufficient generativity to make use of the formal character model. While it might conceivably increase player engagement for them to know their character had had a pet rabbit as a child, it won’t affect the gameplay unless the game’s AI can somehow take advantage of that knowledge.
Alternatively, the system could be used as an adjunct to existing tabletop games, either to help beginning players design their characters, or more likely to help GMs quickly flesh out random NPCs that they need. The GM could use generic questionnaires for merchants, innkeepers, etc., when a given character types was needed. Or, since the system is driven by a SAT solver, the system could generate answers to the questions itself and simply present the finished character model to the GM.
kind of user-interface to allow a player or GM to drive a
Questions, answers, and their implications are specified in a human-authored text file. Questions are prefixed with Q: and answers with A:. Implications for a question or answer, if any, are given in a comma-separated list beneath their respective question or answer. Questions and answers are free-form text. Implications are given as atomic propositions (single tokens) or applications of predicates to terms, expressed in ersatz English. For example: Q: What could you have done to save your brother's life?
dead brother, family_guilt A: Taken the car keys away
alcoholic brother A: Taken him away from dad
abusive father A: Made him move in with me to get him out of the neighborhood
brother_gang_member
propositions and multiword constructions such as dead brother are predicate applications (i.e. dead(brother)).
ified as part of the predicate.
The act of asking the question implies that the character’s brother is indeed dead, and that the character feels some responsibility for it. It therefore forecloses any questions that presuppose the brother is still alive or that the character is an only child. The first answer, if chosen, further adds that the character’s brother was an alcoholic, while the second adds that their father was abusive. The third answer adds that the brother was a gang member.
The questionnaire can also specify a limited set of constraints and other axioms: • • • • •
Mutually exclusive: fact1, … , factn
Contradiction: fact1, … , factn fact1, … , factn cannot all be simultaneously true. Unique: fact1, … , factn
conclusion <- premise1, …, premisen
conclusion <= premise1, …, premisen
& Lifschitz 1988; Gebser et al. 2012). This differs from classical implication in that it adds the constraint that the conclusion may only be true if one of its Horn rules justifies it.
// Afflictions Unique: insomnia, violent, asthma, grief_stricken, ignored, tone_deaf, bored, superstitious, bad_temper
The questionnaire is a more author-friendly front-end to the internal KR language used by the character modeling system. Character models are represented in an order-sorted logic with atomic terms: terms (arguments to predicates, represented internally by strings) are atomic rather arbitrary term expressions; they’re divided into sorts (data types); and those sorts are ordered (types can have subtypes) with Entity being the top sort. Predicates specify sorts for their arguments: dead( ) is limited to ’s of the Person sort.
One of the goals of the system is to allow non-technical
authors to write questions and answers for the system. As a
result, we’ve been reluctant to add rules with explicit
variables and quantifiers, such as would be found in Prolog
If the author writes “dead father”, i.e. dead(father), that has the usual semantics since father is a term. But if they write “dead parent”, since parent is a sort, it means ∃ ∈ parent. dead( ). A question or answer can therefore add to the character model that they have a dead parent, without having to specify which parent. The current language, however, does not have a syntax for specifying that all parents are dead (apart from saying “not living parent” or explicit quantified variables such as ∃ . dead( ) ∧ abusive( ), i.e. they have a dead, abusive parent.
• • • generalizes ∀ ∈ . ( ) → ( ), where is the sort over which is defined. strongly generalizes ∀ ∈ . ( ) → ( ), but also, ( ) implies at least one of the predicates that is strongly generalized by it must be true of . negatively generalizes
Asserts that ∀ ∈ . ( ) → ¬ ( ).
exists. So, if you love or hate someone, they must exist.
ists, so if someone exists, they must also be living or dead.
so you can’t be both living and dead.
cates are currently limited to at most two arguments.
Technologies 2004) game engine. It uses the CatSAT logic programming system for back-end inference. The KR language is translated at run-time into CatSAT assertions and character models are computed by solving for models of the assertions. ( ) <- ( ) ¬ ( ) <- ( ) ( ) <= ( ) ( ’) <= ( )
Before the questionnaire can be administered, the information in the questionnaire must be compiled into propositions and axioms in the SAT problem. Each potential fact in the questionnaire, be it an atomic fact such as “insomnia”, or a predicate instance, such as “likes alcohol” is mapped to a proposition in the SAT problem. The system also generates implication rules for any generalizations or existential quantifications over the predicate instances.
be the set of arguments to that appear as facts in the questionnaire. Then, to implement generalization, for each ∈ , the system adds the rule:
the rule:
adds the rule:
subsort ’ ⊆ , and for each ∈ ’ ∩ , the system adds the rule: CatSAT operates internally on clauses in conjunctive normal form, i.e. disjunctions of literals (propositions or their negations). The system translates standard <- implications into single CNF clauses, i.e. ( ) <- ( ) is translated into the single clause ( ) ∨ ¬ ( ). However, <= rules are more complicated, since they have the semantics that the consequent can only be true if the antecedent of one of its <= rules is true. A set of rules, such as: ( ) <= ( ) ( ) <= ( ) ( ) ∨ ¬ ( ) ( ) ∨ ¬ ( ) ¬ ( ) ∨ ( ) ∨ ( ) is translated first into the biconditional ( ) ↔ ( ) ∨ ( ), which is then translated into its CNF form:
exclusive: are directly supported by CatSAT, and so require no preprocessing other than to map facts in the questionnaire to the internal SAT propositions used to represent them.
assumed facts is the subset of ’s answers ( ), for which the assumptions , ’s implications ( ), and that answer’s implications are all consistent, i.e., they have a model: ( , ) = { ∈ ( ): ∃ . ⊨ ∪ ( ) ∪ ( )}
∪ ( ) ∪ ( ).
The basic loop is then to accumulate a set of facts by asking questions that don’t contradict , adding any implications of those questions and answers to :
} foreach ∈ questionnaire { // Find the non-contradictory answers = ( , ) // Make sure there are enough of them if ⌊ ⌋ > 1 { present and to the user collect user’s answer ∈ // Update the known facts about the character = ∪ ( ) ∪ ( ) }
⊨
To get a sense of the operation of the system, we give an example run using a minimal questionnaire, which can be found in the appendix. “Implications” gives the set of specific implications of the questions and answers (the set from above). “Also inferred” gives other propositions that the SAT solver would determine to be true, but that don’t appear in the set :
Also inferred: not posh tastes, since that contradicts simple tastes.
Also inferred: since the character is inflicted with insomnia and they aren’t allowed multiple afflictions, the other inflictions have been ruled out. Since they have been established as sentimental, other mindsets such as being arrogant have also been ruled out (this is not a realistic personality model, but it’s one of the axioms).
Also inferred: hates father, not brother living, guilt (from family_guilt)
Also inferred: dead father. Note that an option involving the father’s favorite book is suppressed here because the character hates the father.
Also inferred: since there are living family, but the brother and father are dead, it knows there must be another living family member
to make a random character model consistent with the inferences: the character has a mother, whom the character neither loves nor hates; they have a pet; they’re gregarious, extensions are necessary are best determined by putting the frail, and religious. systemin front of users.
We are not aware of any previous work on character PCG guilt <- family_guilt
using questionnaires. However, there is a sizable body of
work on SAT-based PCG in general using Answer-Set Pro- // Personalities
gramming
Therehasalsobeenasignificantamountofworkonmak- Unique: insomnia, violent, asthma,
ingauthor-friendlylanguages for interactive fiction. Ingold grief_stricken, ignored, tone_deaf,
(2015), co-designer of Ink
Nelson argues that within the IF domain, English can be Mutually exclusive: simple tastes, posh
used effectively as a declarative language
// QUESTIONS Q: What'll you have to drink?
A: Whiskey, on the rocks.
Conclusion likes alcohol, simple tastes AutoDread is a work in progress. At this point, we have a A: A bottle of spring water, please workingparser,inferencesystem,anddriverloop. Thenext posh tastes step is to build out a more substantive questionnaire, which A: A diet coke. I'm trying to watch my we’d like to do in conjunction with writers. Our hope is to weight. build a system that technophilic non-programmers can au- health_oriented, likes sweets thor for. Such users are often found in game development, A: An ice-cold bottle of orange juice. IF authoring, and table-top roleplaying. I'm parched
This will certainly involve extending the system. As likes sweets mentioned above, adding new predicates (and hence new verbs) currently requires editing the underlying C# code, in Q: When you have trouble sleeping, what part because it requires giving the parser hints as to how to do you focus on? translate between the internal formof the assertions and the insomnia pseudo-Englishofthequestionnaire. It wouldalsobedesir- A: The last time my brother and I saw able to find natural Englishexpressions for the higher-order each other assertions about predicates, such as generalization. How- brother, sentimental ever, it’s unfair, or at least unrealistic, to ask naïve users to A: The time I won a big game back in learn the difference between the material implication of high school classical logic and Horn clauses with stable-model seman- athletic, nostalgic tics, both of which are supported in the system. A: What I'll say to my lover when I
It may also be necessary to extend the expressiveness of make it back the system’s KR language. However, what types of optimistic, lover
A: The serenity of mountain climbing likes outdoors, peaceful, athletic Q: What do you have in your pockets? A: My inhaler. I'm not in the best of health.
asthma, frail A: My trusty-dusty pocket knife. You never know when something (or someone) will need cutting.
prepared, violent A: A battered paperback novel. Rule number two: always have something to read.
prepared, likes literature A: My lucky coin.
Superstitious Q: What do you miss the most about the before times?
nostalgic A: All the people. It's lonely in the wastelands.
sentimental, gregarious A: There used to be a lot more to do around here. Everything is so boring these days.
playful, bored A: My family. They were all killed in the incident.
mother dead, father dead, likes mother, likes father, nostalgic, grief_stricken A: I had a dog. Now, I have nothing.
pet dead, grief_stricken Q: What do you think that you're better at than you really are?
arrogant A: I like to think I'm pretty funny. No one else seems to agree.
obnoxious A: I'm a really good singer! Probably.
tone_deaf A: I can weather any storm. As long as it's not a metaphor for a difficult emotional experience.
vulnerable A: I'm a talented leader; just, most of the time, people ignore my guidance.
bossy, ignored
Q: What could you have done to save your brother's life?
dead brother, family_guilt A: Taken the car keys away
alcoholic brother A: Taken him away from dad
abusive father A: Made him move in with me to get him out of the neighborhood
brother_gang_member Q: What book do you read every year on the anniversary of your father's death?
dead father A: The bible
religious A: "Ender's Game". He loved it.
loves father A: King Lear.
narcissistic father Q: Why are you the black sheep of the family?
black_sheep A: I married outside of our faith
religious_family A: I just have this temper
bad_temper A: I like the bottle too much
alcoholic self A: I'm a fucking loser
hates self Q: How often do you see your family?
living family A: Once a year A: Once a week
loves family A: Never estranged_from_family