Failing Quickly

This week, Rae and I explore our Gossip 1.1 Method and determine its strengths / weaknesses. Additionally, I continue my reviewing of the literature for gossip systems.

Research Through Design

So for this week, Rae and I began exploring using the technique of Subject, Action, Target based on the RDF triple techniques described by Mooney and Allbeck (2014) as well as the category based system described in Brown et al (2021). Below, the details of our system are described. Then, after going through and implementing the solution, we evaluated how well it worked regarding several different criteria.

Finally, we then went through and revised the system for us to begin prototyping again for this coming week.

Implementation of the System

For our first version of our new Gossip System, we were primarily aiming to create a system that would:

1. Allow NPCs to have their own views / knowledge

2. Allow NPCs to witness actions taken by the player to inform their knowledge

3. Allow NPCs to gossip about their knowledge in a story like format that matches their character

4. Allow NPCs to interpret the knowledge shared in different ways

5. Make this process noticeable to a player

These goals were formed from our readings of Kreminski (2023) and Mooney and Albeck (2014). To accomplish this goal we decided we would create a new method of sharing information in which NPCs would observe who others spoke to, went, and said. Then, using this information, they would then post on our games equivalent of a social media site which the player could view. Finally, other NPCs would be able to view these posts (informing their knowledge) and then post their own replies in a way that matches how they speak (allowing for different interpretations).

To create the ability to make these posts, we decided that each gossip would include the format of {subject} {action} {target} such as {Alex} {kissed} {Tyler}.

Through doing so, we would then be able to create posts such as

Someone just told me {subject} {action} {target} but i don't see what the big deal is

Which would then be translated to 

Someone just told me Alex kissed Tyler but i don’t see what the big deal is

Then we decided to use a tagging system where we would give each of the different gossips a set of labels such as {Embarrassing} and {Dog} to represent that the information is embarrassing and the topic involved dogs. We decided to create both a positive and negative tag category so that it would inform what the attitude towards the topics was for the gossip. Such as pro dog or anti dog. From there, we created a set of generic replies for each character that they would use for different gossip based on their tags. For example

Name	Positive Tags	Negative Tags	Reply
Olive	Cool, Surprising		EYYYYY
Olive	Surprising	Surprising	Thank you for bringing us the HOT DETAILS
Olive		Inappropriate	MAN....that's disappointing...{subject}....

Finally, we create a simple text based method to try different combinations of NPCs making posts for various pieces of gossip and seeing what replies would be generated. 

Evaluation of the System

After going through the process and creating these different systems, the following conclusions were drawn:

1. Unused / Unclear information: The positive vs negative tag system mainly resulted in more headaches rather than benefits when trying to come up with different replies. Additionally, it was unclear at times what 

2. Loss of in person communication: Currently the system would just share information as soon as the NPC receives it and then any other NPC would respond if able to. However, we discussed and felt that it was lacking a lot of potential dynamics such as seeing characters react long form and seeing the NPCs interact with each other.

3. Issues with Actions: The action / target separation ended up having little value since we found very few situations to use those elements separately. Additionally, the current list of actions were found to be VERY generic resulting in the gossip losing a lot of specificity making it feel less interesting. Finally, when trying to use the actions, we found issues with some replies needing it in present tense and some needing it in past tense.

New Implementation Details 

Based off the conclusions we drew from this last iteration, we have gone through and simplified our system to the following:

• Gossip Data

  • Subject: Who did it

  • Event: What they did

    • Past-tense plain text

    • Present-tense plain text

  • Tags: What type of thing did they do?

These solve the above three issues by

1. Simplifying the tag system

2. Simplifying the Action / Target System into an event system

3. Allowing for texts that have some using past and some using present tense

Additionally, we have added a new feature in which some gossip will only be shared in person initially matching the human behavior of wanting to not spread gossip too fast but get the most bang for their buck (Crawford 2004). This then allows us to now introduce a new type of interaction where instead of always walking up to conversations to participate, you will also have opportunities to instead go to a location to ease drop on people talking.

From here, we will take this next week to go through and implement each of these changes.

Literature Review

Game Design Research: An Overview

This overview serves as the introduction to the book Game Design Research by Lankoski and Holopainen. In their overview, they provide an introduction to what game design research is and the differing ways people have gone through conducting it. They describe game research as:

The aim is to uncover new facts and insight about game design, design processes, or games as designed objects; that is, to gain new knowledge and understanding about game design. (Lankoski and Holopainen 2018)

They then list the following methods of conducting game design research:

1. Autobiographical Approaches

   1. Think like postmortems / journaling and thinking on process

2. Comparative and Conceptual Approaches

   1. Think like trying to define what game design and game systems are

3. Research Through Design

   1. Think like having a goal and then creating a game to explore that goal like “how do we create a framework for implementing a gossip system”

4. Design and Evaluation Methods Research

   1. Think like how can we go about making a game and then determining if that game is a good game?

5. Studio and Developers at Work Studies

   1. Think like asking developers about their processes or analyzing how they work through observations to draw conclusions on the process

Additionally, they also provide throughout various researchers work for reference material regarding what research through design looks like.

Overall, they provide a very good formal explanation of games research which is helpful considering that is what I am trying to accomplish. Currently, I would say that my research primarily lies in (3) with my gossip system and (4) with my believability metric.

Grapevine: A Gossip Generation System

In Khosmood and Walker’s paper, they describe their own approach to how to generate a gossip conversation between two characters. Rather than looking at how gossip can be spread, their efforts were more so focused on how you could create dialogue using generative techniques. In particular they focused on using character personality traits based on the Big Five that would inform what a character should say based on what they know. From there, they then developed a system that would structure conversations as

Outreach -> (Deflection -> Outreach 2 -> Outreach 3)? -> Tease A -> Tease L -> Tease R -> Reveal -> Reaction

Additionally, they also had each character have both their own knowledge as well as their own internal understanding of what others likely knew with a floating based scale from 0-1 representing confidence. 

Finally, they concluded their paper with results from a 50 person survey which showed that people were able to very effectively determine what the conversation was about but only picked up on a few traits the character had. However, the metric was called into question by the researchers so the validity of these results is suspect.

Overall, this approach will likely be used as a starting point as we begin to look into how to have engaging conversations to eavesdrop on. While we still may lean towards prewritten narratives for those, the overall flow of a gossip conversation may still be of use.

Interactive Storytelling: Chapter 14 HistoryBooks and Gossip

In this chapter, Crawford discusses how he went about implementing his own gossip system in a post mortem style of research. Essentially, the key takeaways were:

• HistoryBooks: To store information regarding what an NPC knows, its beneficial to have a method to keep track of the events they experience. For each event that occurs, mark which NPC witness it, and what the prior event that caused it were, if its secret, if it's a lie, the time, and the location

• Gossip: To handle gossip, each NPC should keep a list of headlines that it draws upon when needing to share and then creates a tale around those headlines by selecting the prior information that occurred in the same period of time or location and then filtering it to the relevant bits. 

• Who to share with: You determine based off the following equation and who does not know the information already

  • Value of telling Event A to ActorB = Import[Verb[Event A]] ×  ➥(AbsoluteValue(PerVirtue[ActorB, Subject] × PerVirtue  ➥[ActorB, DirObject]) + AbsoluteValue(PerIntegrity[ActorB, Subject] ×  ➥PerIntegrity[ActorB, DirObject]) + AbsoluteValue  ➥(PerPower[ActorB, Subject] × PerPower[ActorB, DirObject])))

    • Crawford (2024)

• Additional Features: You can also consider having NPCs lie to one another in order to accomplish a goal, trying to keep secrets that spread for dramatic reasons, and grapevines to trace who told who for NPCs to hunt down who leaked or lied.

  
  

However, Crawford also brings up some gaps he has found in his approach

• How to handle partial knowledge of a Tale in a way that's dramatic but not chaos to the PC

• How to handle NPCs knowing only part of the details of an event and catching up on the details

• How to handle the tale that describes after the headline occurred

  
  

Overall, Crawford presents another interesting way of handling the tracking of gossip information which may be a viable approach for handling the overhearing situation. However, Crawford leans on tracking sequences of events where as for Rae and I, we are tracking individual moments.

The Wrap Up

All in all, not a bad week of development. Looking towards next week, Rae and I should hopefully have some promising developments to discuss.

Work Cited

Brown, J., Lee, J., & Kraev, N. (2021). Reputation Systems for Non-Player Character Interactions Based on Player Actions. Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, 13(1), 151–157. https://doi.org/10.1609/aiide.v13i1.12950

Crawford, C. (2004). Chris Crawford on interactive storytelling. New Riders Games.

Khosmood, F., & Walker, M. (2010). Grapevine: A gossip generation system. Proceedings of the Fifth International Conference on the Foundations of Digital Games, 92–99. https://doi.org/10.1145/1822348.1822361

Kreminski, M. (2023). Toward Better Gossip Simulation in Emergent Narrative Systems. 2023 IEEE Conference on Games (CoG), 1–4. https://doi.org/10.1109/CoG57401.2023.10333140

Lankoski, P., & Holopainen, J. (2018). Game design research: An introduction to theory & practice. ETC Press.

Mooney, J., & Allbeck, J. M. (2014). Rethinking NPC Intelligence—A New Reputation System. MIG ’14: Proceedings of the Seventh International Conference on Motion in Games, 55–60. https://doi.org/10.1145/2668084.2668091