Keywords

1 Introduction

In the game Tyranny, the player is often presented with choices where the values of at least two opposing parties are at stake [1]. For example, after a successful siege of a city, the player may let one group, the Scarlet Chorus, loose so that they may raid and pillage as they please. On the other hand, the player may send in another group, the Disfavored, to impose martial law. The factions are representatives of sets of opposing values. The Scarlet Chorus represents, at least, hedonism, while the Disfavored represents conformity. In interactive narrative, this situation is a dilemma, or a moment wherein a player is presented with alternatives that each harm at least one value.

In the interactive narrative field, dilemmas have been utilized in order to generate stories that maintain some level of narrative interest by presenting players with hard choices [2, 3, 7]. In research and in video games, dilemmas are presented via non-player characters (NPCs) who represent the values at stake. We call this kind of character a value-sensitive agent [3].

Value-sensitive agents present a coherence problem. Coherence between the actions available to the player in a dilemma, the values they harm or promote, and the attitudes of the characters presenting the dilemmas through which the player interprets the alternatives is maintained solely by the author. With the current state of the art, if an author were to try to replicate Tyranny, they might accidentally write a situation wherein a character from the Scarlet Chorus acts disgusted when a player chooses to act hedonistically. Without extra writing to contextualize this reaction, this character appears to be acting inconsistently.

Fig. 1.
figure 1

On each round, the player is presented with a dilemma (a). After a decision is made, the advisors react (b). At the end of the game, the advisors decide whether to retain the player as emperor by majority vote (c).

Full size image

Toward reducing the possibility of inconsistent NPC behavior, we introduce a method of value-sensitive agent generation that enforces relationships between the values such that no values that are diametrically opposed can be cherished at once by the same agent. The values and the relationships between them are given by Schwartz’s Theory of Basic Values [8]. We contextualize our implementation in our role-playing game, Carambola. Here, the player takes the role of an emperor making executive decisions, and is presented a sequence of dilemmas that lead either to them retaining their throne or losing it. Each choice the player makes elicits reactions from their NPC advisors based on the advisors’ attitudes toward the values that the actions affect. Their reactions nudge them toward voting for either one of the player’s possible outcomes (see Fig. 1). We hypothesize that our design improves the advisors’ believability [4].

2 Theoretical Background

At the core of Carambola’s design is an implementation of Schwartz’s Theory of Basic Values (TBV), which posits that there is a shared set of values across cultures worldwide [8]. The theory has two facets which are relevant to our implementation: the values themselves and a geometric model for how the values are related to each other. According to the TBV, these universal values can be placed in a circular continuum in which the proximity of the values represents the amount of similarity their underlying motivations have (see Fig. 2).

Fig. 2.
figure 2

The continuum of values introduced by Schwartz’s Theory of Basic Values. The proximity of the values reflects how similar their underlying motivations are.

Full size image

In Carambola, we use this adjacency relationship from the TBV to enforce restrictions on the possible configurations of advisor attitudes toward the values. The advisors are each motivated to promote or maintain their empire’s general wellness. However, that motivation manifests differently for each advisor through their personal attitudes toward the values that are promoted and harmed by the player’s choices. Following the EGAD framework, the advisors can hold three attitudes: cherish, despise, or ambivalent about [5]. We extend the framework by ensuring that the values that advisors cherish (or despise) lie adjacent to each other on the continuum.

3 Related Work

Carambola’s design is inspired by past interactive projects which feature value-sensitive agents [6, 10]. Behind one of these projects [10] is one of the earliest examples of value-sensitive narrative generation, IDTension [9]. Here, an author defines values via abstract keywords (e.g., non-violence, law). The author also configures both agents’ attitudes toward each of these values and actions that are symbolic of the values. Carambola takes a similar approach to agent and action design. However, in Carambola the values are not defined by us, the designers, but are rather come directly from the TBV. Furthermore, our implementation enforces relationships between the values, precluding authorial mistakes that can produce situations in which agents can simultaneously hold positive attitudes toward diametrically opposed values, such as violence and non-violence.

The main mechanic of our game, dilemma resolution, is largely inspired by proposed frameworks for dilemma generation. To inform our NPC design, we follow the EGAD framework [5], which was proposed as a refinement over the GADIN system [2]. In this framework, agents may cherish, despise, or be ambivalent to each of the values, while actions may promote, harm, or do nothing to each. Like EGAD, we use the values from the TBV. However, the framework ignores the embedded relationships between the TBV’s values. In Carambola, these relationships are central to the NPC design.

4 Game Design

To facilitate the player’s ability to make reasoned decisions in the game, we sought to construct the advisors so that what their reactions and reasoning are consistent and clear to the player. We achieve consistency by ensuring that the values that the advisors cherish (or despise) are adjacent to each other on the value continuum. Clarity is achieved by plainly stating the advisors’ reactions, their affinities toward the player, and by labeling actions such that there is a simple thematic link between what they represent and the values they promote.

4.1 Action Design

On each round, Carambola presents the player with a dilemma of two alternative actions (see Table 1). The action specifications were handwritten so that their effects fit in thematically with their labels. For example, Authorize Military March shows off the glory of Carambola’s military (promoting their achievements) while reinforcing the force that the empire has over its citizens (harming universalism). Upon choosing an action, the player triggers its effects on the values, which elicit reactions from the advisors.

Table 1. A list of all available actions Carambola, along with their effects on the values. Every action has an opposite version, where its effects are flipped from the original.
Full size table

4.2 Advisor Design

To facilitate discussion about our advisor design, we will refer to Dmitri, an example advisor that can be generated in the game (see Fig. 3).

Value Attitudes. Like dilemma generation systems in the past, we designed advisors so that they each have values that they cherish, despise or are ambivalent toward [2] [3] [5]. At game start, we generate their attitudes toward the values according to the following rules:

  1. 1.

    The two values that an advisor cherishes (despises) are adjacent on the value continuum.

  2. 2.

    For an advisor to despise a value, it must be on the opposite half of the value continuum from the ones that they cherish.

Consistent with the TBV, with these two rules, adjacent values have similar motivations, while values that are opposite from each other on the continuum can conflict [8]. In our example, Dmitri cherishes power and security, but not power and universalism because the latter pair are on opposite sides of the continuum. For the purpose of ensuring that Dmitri’s values are clear to the player, this is ideal: while power and security together emphasize control and the overcoming of threats, universalism invites diversity and self-expression.

Fig. 3.
figure 3

Value attitude configuration for Dmitri, an example advisor. Dmitri cherishes power and security, and despises self-direction and universalism. He is ambivalent toward the rest of the values.

Full size image
Table 2. An illustration of how the effects (promote or harm) of the player’s choice and an advisor’s attitude (cherish or despise) interact to produce points.
Full size table

Reactions. After the player chooses one of the alternatives presented to them, each of their advisors takes a turn to react. Computationally, an advisor’s reaction is the sum of points that the player’s choice yields, with points being given according to Table 2. Thus, an advisor’s reaction can range from being very positive (yielding 2 points) to neutral (yielding 0 points) to very negative (yielding −2 points). This sum is added to the advisor’s overall affinity to the player.

To illustrate, suppose the player chooses Maintain Barracks. Because Dmitri cherishes power and this choice promotes it, the player receives +1 point. Because Dmitri despises universalism and this choice harms it, the player receives +1 additional point. Dmitri’s overall reaction, then, is very positive, giving +2 points. His affinity toward the player moves in the positive direction.

5 Future Work

5.1 Evaluation of the Character Model

We plan on evaluating via user study the effect that Carambola’s implementation of the TBV has on the believability of its NPCs. Our hypothesis is that our implementation makes the agents more believable than if their attitudes toward the values were randomized. To assess this, we will extend the list of actions in Carambola so that there are pairs of actions that affect the same values in the same way. For example, we may introduce the following two pairs:

  1. 1.

    Increase Weapons Manufacturing and Occupy a Neighboring City, which each promote power and harm universalism.

  2. 2.

    Enforce Attendance to Mass and Close All Business for Holiday, which each promote tradition and harm self-direction.

Our evaluation will have two study cases that differ in how the advisors would react to player choices. Our test case will be of the implementation detailed thus far, where advisors react according to their values. For each action pair, they will react exactly the same way for either action. In the control case, an advisor will favor one action in a pair, but disfavor the other one. Intuitively, if our hypothesis is correct, the control case would introduce a level of inconsistency in the advisors’ reactions that will break players’ suspension of disbelief.

To test our hypothesis, we will use the methods proposed in Gomes et al. [2013] to quantify the difference in character believability between the control and test cases [4]. Believability is split into a number of dimensions that describe different aspects of agent behavior. For example, one dimension is behavior coherence, which is the degree to which a human observer may deem an agent’s actions to be logical according to their mental model of the agent’s state. Part of our future work will be to find which subset of dimensions are appropriate for the limited context of our game.

5.2 Extending the Character Model

Currently, the advisors’ attitudes toward the values are of a ternary set: cherish, despise, or ambivalent toward. This causes two issues: the advisors cannot hold positive attitudes toward conflicting values, and the advisors can only care about four of the values at a time. Even in Carambola, one can imagine space for characters who hold more complex attitudes. For example, a Robin Hood-esque character might seek more power in order to spread benevolence among Carambola’s peasants. We speculate that switching from a discrete set of attitudes to a continuous one and allowing agents to have attitudes for all values on the continuum will be a step toward introducing such nuance.

5.3 Refining the User Interface

One relevant question to both the player experience and our described evaluation is that of how the presentation of each alternative action affects a player’s ability to understand why their decisions elicit certain reactions from the NPCs. In Carambola’s current iteration, we hand-wrote labels for each action (see Table 1). Whether the values that are promoted or harmed are appropriate for each action is highly subjective. Disagreement between the authors’ and players’ mappings between actions and values may weaken the data we gather in our evaluation.