Welcome, Guest. Please login or register.

Login with username, password and session length

Advanced search

1412038 Posts in 69446 Topics- by 58482 Members - Latest Member: ZerusW

June 21, 2024, 03:23:20 AM

Need hosting? Check out Digital Ocean
(more details in this thread)
TIGSource ForumsDeveloperTechnical (Moderator: ThemsAllTook)Procedural resource dump
Pages: 1 ... 23 24 [25] 26 27 ... 30
Author Topic: Procedural resource dump  (Read 142624 times)
Level 10

The archivest master, leader of all documents

View Profile
« Reply #480 on: February 12, 2023, 07:39:00 PM »

 Let's build GPT: from scratch, in code, spelled out.

00:00:00 intro: ChatGPT, Transformers, nanoGPT, Shakespeare

baseline language modeling, code setup

00:07:52 reading and exploring the data
00:09:28 tokenization, train/val split
00:14:27 data loader: batches of chunks of data
00:22:11 simplest baseline: bigram language model, loss, generation
00:34:53 training the bigram model
00:38:00 port our code to a script

Building the "self-attention"
00:42:13 version 1: averaging past context with for loops, the weakest form of aggregation
00:47:11 the trick in self-attention: matrix multiply as weighted aggregation
00:51:54 version 2: using matrix multiply
00:54:42 version 3: adding softmax
00:58:26 minor code cleanup
01:00:18 positional encoding

01:02:00 THE CRUX OF THE VIDEO: version 4: self-attention
01:11:38 note 1: attention as communication
01:12:46 note 2: attention has no notion of space, operates over sets
01:13:40 note 3: there is no communication across batch dimension
01:14:14 note 4: encoder blocks vs. decoder blocks
01:15:39 note 5: attention vs. self-attention vs. cross-attention
01:16:56 note 6: "scaled" self-attention. why divide by sqrt(head_size)

Building the Transformer
01:19:11 inserting a single self-attention block to our network
01:21:59 multi-headed self-attention
01:24:25 feedforward layers of transformer block
01:26:48 residual connections
01:32:51 layernorm (and its relationship to our previous batchnorm)
01:37:49 scaling up the model! creating a few variables. adding dropout

Notes on Transformer
01:42:39 encoder vs. decoder vs. both (?) Transformers
01:46:22 super quick walkthrough of nanoGPT, batched multi-headed self-attention
01:48:53 back to ChatGPT, GPT-3, pretraining vs. finetuning, RLHF

01:54:32 conclusions

00:57:00 Oops "tokens from the future cannot communicate", not "past". Sorry! Smiley
01:20:05 Oops I should be using the head_size for the normalization, not C
« Last Edit: February 12, 2023, 07:44:53 PM by gimymblert » Logged

Level 10

The archivest master, leader of all documents

View Profile
« Reply #481 on: February 12, 2023, 08:09:13 PM »

Training strategy

 How ChatGPT is Trained

The Illustrated Retrieval Transformer
Summary: The latest batch of language models can be much smaller yet achieve GPT-3 like performance by being able to query a database or search the web for information. A key indication is that building larger and larger models is not the only way to improve performance.

 New ChatGPT Competitor Explained Claude vs ChatGPT

 InstructPix2Pix Edit Image using Text Instruction High Level Explanation Hugging Face Space Demo

 Beyond Text - Giving Stable Diffusion New Abilities
 Stable Diffusion & Friends: High-Resolution Image Synthesis via Two-Stage Generative Models
 Food for Diffusion

« Last Edit: February 12, 2023, 09:31:13 PM by gimymblert » Logged

Level 10

The archivest master, leader of all documents

View Profile
« Reply #482 on: February 12, 2023, 08:12:30 PM »

 ChatGPT with Rob Miles - Computerphile

Do As I Can, Not As I Say:
Grounding Language in Robotic Affordances

 Supplementary video for Do As I Can, Not As I Say: Grounding Language in Robotic Affordances

 Supplementary video for Inner Monologue: Embodied Reasoning through Planning with Language Models

Improving AI Reasoning Skills through Self-Generated Prompts - Why ChatGPT Does Not Have an IQ of 83
 Language Models Perform Reasoning via Chain of Thought

One example of generalized learning is GPT-2's ability to perform machine translation between French and English, for which task GPT-2's performance was assessed using WMT-14 translation tasks. GPT-2's training corpus included virtually no French text; non-English text was deliberately removed while cleaning the dataset prior to training, and as a consequence, only 10MB of French of the remaining 40,000MB was available for the model to learn from (mostly from foreign-language quotations in English posts and articles).[8] Despite this, GPT-2 achieved 5 BLEU on the WMT-14 English-to-French test set (slightly below the score of a translation via word-for-word substitution). It was also able to outperform several contemporary (2017) unsupervised machine translation baselines on the French-to-English test set, where GPT-2 achieved 11.5 BLEU. This remained below the highest-performing contemporary unsupervised approach (2019), which had achieved 33.5 BLEU.[8] However, other models used large amounts of French text to achieve these results; GPT-2 was estimated to have used a monolingual French corpus approximately 1/500 the size of comparable approaches.

« Last Edit: February 12, 2023, 10:22:52 PM by gimymblert » Logged

Level 10

The archivest master, leader of all documents

View Profile
« Reply #483 on: February 12, 2023, 09:33:19 PM »

 New Embedding Model by OpenAI - Intro and Explanation

 OpenAI Embeddings (and Controversy?!)


Level 10

The archivest master, leader of all documents

View Profile
« Reply #484 on: February 12, 2023, 09:51:44 PM »

ai alternative to open ai, some open source and free


Level 10

The archivest master, leader of all documents

View Profile
« Reply #485 on: February 13, 2023, 05:10:52 PM »

I realized I never shared Emily short's groundbreaking work on NPC dialog here, let me document that quickly

Emily's works:

- the basics
http://www.tigsource.com/2009/05/14/emily-short-conversation-methodologies/ (masterclass in the comment)

- series on modeling conversation, cover all the state of the art, practical

- the breakthrough design masterpiece on cnversation, to read!


- less filtered version of the conversation modelling series

- More overkill details about versu

- her analysis of how information, story and design merge in some game is insightful for building our own implementation in any games.

- random

There is also correlation between writing conversation and a game based on scattering information on a spatial basis (and the gating that generally goes with like in adventure games). Instead of mapping the discourse structure into a conversation interface, where the player navigate through a network (generally a tree) of information, the world space became the discourse map by scattering information onto place that act like topics. I do not know if this part is clear Huh?

Homer in silicon is archive here

more Emily
Choice and Character: Lessons from Writing Multiplayer Narrative Games

Turning Comedy of Manners into Gameplay: Versu Postmortem

Beyond Procedural Horizons: Exploring Different Uses of Procedural Content Generation (slide)

MORE about old state of the art
That NPC stuff looks complicated!

However I found that for dialog pile Who, Me? ignore  if not needed but it's the chatterbot perspective, I think there is a few concept explain here that could make it easier for the conceptualization and organisation of the speech data. Not that it's ground breaking (unlike versu) but seeing actual implementation would make your own surely easier.
http://www.gamasutra.com/view/feature/6305/beyond_fa%C3%A7ade_pattern_matching_.php?print=1 (especially the suzette part)
I know It's NLP and that it is overkill for your project, but I thought some concept might still help in how they hadnle topic and various discourse handling. More like for inspiration and thought (mostly for procedural generation parts).

Optional read

Also To add to the thought of the social aspects of my previous and explain the reasoning. The goal was to offer a basic gameplay loop to ease the player into the world as a progression mechanics, a short term goal loop with clear gating of the main currency (information and agency). Other game use trading or fighting, I thought a social ladder mechanics based on intel could be that and ease the player into the the broader hi level goals or exploration without getting lost at first, it's like a mechanical short term breadcrumb.

It's true they originate from a design which I'm working on and deal with more intimate interactions.

So while you don't have to be that intimate, you can substitute individual with group or more abstract concept, simplify the loop or adapt it to your purpose. But I want to raise the necessity for clear short term loop.
Beyond Facade pattern matching can be found here: https://web.archive.org/web/20110607071906/http://www.gamasutra.com/view/feature/6305/beyond_fa%C3%A7ade_pattern_matching_.php?print=1


Keith Stuart on AI, acting and the weird future of open-world games
Why, AI? (2015)

I gonna start caching the data before it get lost to time.

The original ELIZA source code found

 [ML News] Anthropic raises $124M, ML execs clueless, collusion rings, ELIZA source discovered & more
« Last Edit: February 13, 2023, 05:52:08 PM by gimymblert » Logged

Level 10

The archivest master, leader of all documents

View Profile
« Reply #486 on: February 14, 2023, 10:04:18 PM »

https://www.gdcvault.com/play/1015528/AI-driven-Dynamic-Dialog-through (video)
https://www.gdcvault.com/play/1015317/AI-driven-Dynamic-Dialog-through (slide)
https://www.gdcvault.com/play/1015946/AI-driven-Dynamic-Dialog-through (audio)
AI-driven Dynamic Dialog through Fuzzy Pattern Matching. Empower Your Writers!

Analysis: Conversation Design In Games


Level 10

The archivest master, leader of all documents

View Profile
« Reply #487 on: February 15, 2023, 07:40:17 PM »

Story vs. Game: The Battle of Interactive Fiction
A Talk Given at the Computer Game Developer's Convention 1989


Doug Peers into the Future, this graph reveals the next 41 years of IF history. Horizontally I give you Time in years, vertically ART measured in milliprousts, or one thousandths of the qualitative narrative output of Mr. Proust, who wrote some good books. The era that most concerns me is that between now, accurately pinpointed at 1989 I believe, and the spot marked Doug retires, in about 30 gloriously successful years.

The two quantities I track over time are the maximum narrative quality, measured in milliprousts, of the best work of Interactive fiction in each year using each of two approaches - IF STORYMAKING, making interactive stories through hand plotting, scriptwriting, and crafty use of interaction. The other approach, IF STORY GENERATION, generating interactive stories through algorithms and AI.

The term Story Making, as opposed to Story Telling, I owe to Brian Moriarty. There has been quite a passionate and entertaining exchange concerning these two approaches on the Journal of Computer Game Design BBS. I recommend it.

You can see how I'm betting, that during my working years the best results will come from the STORYMAKING approach. Again, in my unhumble opinion, this is because of the dialog gap, the AI gap. After I'm put out to pasture, though, I do see a reversal, when computers first equal, and then surpass, human intelligence. At that point of course all bets are off, but I would say there's a good chance that some excellent games could come out of collaborations with true artificial intelligences.

It's nice to finally have a reliable milestone for computer-human equivalence. I owe this to Hans Moravec and his fun, radical little book Mind Children, which I recommend. Fasten your seatbelts when you read it.

 Undecided Screamy Big Laff Shrug
« Last Edit: February 15, 2023, 07:46:25 PM by gimymblert » Logged

Level 10

The archivest master, leader of all documents

View Profile
« Reply #488 on: February 20, 2023, 12:05:38 AM »

 Narrative Sorcery: Coherent Storytelling in an Open World

 Heaven's Vault: Creating a Dynamic Detective Story

 The Burden of Proof: Narrative Deduction Mechanics for Detective Games

GDC 2015: Jon Ingold - "Adventures in Text: Innovating in Interactive Fiction"

Ink is inkle's scripting language for writing interactive narrative, both for text-centric games as well as more graphical games that contain highly branching stories. It's designed to be easy to learn, but with powerful enough features to allow an advanced level of structuring.
Overboard! Conversation ink example

Procedural storytelling is exploding the possibilities of video game narratives

Wildermyth tells incredible stories that hit surprisingly hard

An absurdist baseball game is tearing up the internet, thanks to fans

Procedural generation
A mostly javascript-centric resource / links list on procedural content generation (PCG).
JavaScript modules
« Last Edit: February 20, 2023, 12:27:17 AM by gimymblert » Logged

Level 10

The archivest master, leader of all documents

View Profile
« Reply #489 on: February 22, 2023, 10:21:00 PM »

Driving Emotionally Expressive NPC Animations and Behaviors with a Designer-Friendly Pipeline

Level 10

The archivest master, leader of all documents

View Profile
« Reply #490 on: February 24, 2023, 01:21:47 PM »

Conversation as Gameplay (Talk)

 ICCC 2015 - Day Three

Can AI tell a good story?

 Gamelab Barcelona 2017 - Emily Short - The uncanny mirror seeing ourselves in AI


Beyond Branching: Quality-Based, Salience-Based, and Waypoint Narrative Structures

ICIDS: The future of interactive storytelling, plus some Versu thoughts

Theme and Interaction: Designing Action and Puzzles Relevant to Story
    What interaction adds to narrative
        Resolving Challenges (mostly puzzles in IF)
    Deciding what sorts of action are appropriate within the story
        Some standard action types in existing IF
            Pure exploration
            Manipulating machinery and/or physically realistic systems
            Magic use
            Detective work and discovery
            Word-play and surreal interaction
            Conversation/emotional interaction
    Designing a model supporting this action type
        Determining the essential elements of the model; things and kinds
        Considering how the elements interact; properties, adjectives, and relations
        Displaying the relation of model elements to the player; inventory listing, room description, printing the name, and more
        Creating actions whereby the player can manipulate the model

Setting — see also J. Robinson Wheeler’s article

    Creating rooms and objects (Chapter 3 of Inform Manual)
        Place as a reflection of its history and purpose in the story
        The gun over the mantlepiece rule more important than ever, since you must physically supply any objects that are going to be critical in the action
        Purposes of room descriptions in communicating possibilities to the player

    Varying descriptions of rooms
        Using embedded [] elements in say phrases and descriptions
        Using “writing a paragraph about”, etc., to control the way objects are described
            Discussion of whether it’s necessary to report everything present in a room; drawing the player’s attention to one thing or another. (see Nothing More, Nothing Less)

    Designing a map
        Techniques in map design for creating a coherent geography, etc.
        Backdrops to create scenery visible from multiple places (Chapter 3)
        Doors, vehicles (Chapter 3)
        Tying places together with descriptions of travel between rooms (“Up and Up”)
        Large locations represented by multiple rooms (“Tiny Garden”, “Stately Garden”)

    Interactive parts of a setting
        Machinery, e.g., that reveals something about what the area is used for
            Devices, on/off switches, settings and so on; creating mechanisms
            Adjectives and physical states
            Actions that apply to numbers and units (chapters on Units, understanding)

Atmosphere — making areas seem alive and setting a mood (see Anchorhead)

    Every turn rules
        Evoking sounds, smells, etc. in a given location (Atmospheric Effects extension)
        Making crowds seem active
        Giving even quiet non-player characters something to do
        Using tables or extensions (Text Variations, List Control) to produce randomized text

        Creating a schedule of background events (see chapter on Time)
            Weather, time of day (Weather extension)
            Shops opening & closing, etc. (“IPA”)

Exposition — letting the player discover material without info-dumping

    Writing a good introduction — see “The Overture” in Craft of Adventure
        When play begins: … rule
    Hinting, drawing the player’s attention in the right directions
    In-game sources of information
        Books, computers, consultable devices (Basic Actions chapter, “Reading and Talking”)
        Other characters (though postpone detailed discussion of conversation)


    Planning scene structure
        Using the scene mechanism (Chapter on Scenes)
        Moving the player from place to place (chapter on Change)
        Changing rooms and moving objects around between one scene and the next (chapter on Change)
    Using time out of sequence
        Flashbacks (see All Hope Abandon, “Pine”)
        Temporally non-linear narrative (see Photopia)
    Linear vs. branching narrative
        Using the scene mechanism to handle choice alternatives by the player
            Presenting choices clearly; omitting irrelevant choices. See Chris Crawford on Interactive Storytelling
            Function of choice in interactive story-telling
                Moral decision-making (see Stephen Bond, Victor Gijsbers; Fate, Floatpoint, Slouching Towards Bedlam)
                Characterization of the viewpoint character
                (Probably some others I’m not really thinking of at the moment.)

    Conveying the shape of the narrative
            In descriptions and events, much as in static fiction
            In action, by teaching the player commands that are going to become important later
                Teaching complicated new actions in small steps
                Giving simple initial challenges to the player
                Creating actions that work in a consistent way throughout the IF
                    Designing a new action
                    Designing a model to support it
        Tracking and signalling progress through the story
            Score and alternatives to scoring
                Score manipulation commands (Chapter on Time)
                The-story-so-far summaries and the like; using tables to build these; “Goat-Cheese and Sage Chicken”
            Using the status bar (make analogy to page numbers, if writers are resistant to having one)
                Status line elements — “changing the status line”, “rule for constructing the status line”
                Basic Screen Effects
            Concept of “plot progress” or triggers — what the player knows or has done so far, or important things that have happened, relative to what still remains to happen
    Writing endings
        Single and multiple endings
            Function of multiple endings
            The function of losing endings in the work overall: dead ends and failure states
        “when play ends” rules

Pacing: individual scenes

    Actions as they affect perceived pace for the player
        Types of action (see Attack of the Yeti Robot Zombies)
            Low-suspense commands like LOOK vs. more dramatic actions
            Density of dramatic behavior –> perceived intensity of scene
                Overview of the standard library actions and which produce dramatic results; using the ACTIONS index
                Controlling the passage of time; making actions that take no game time (“Instant Examine and Look”, etc.)
        Quantity and quality of feedback
            Replies different from the standard library response (see also below)
            Giving enough response to an action that a player remains interested but not so much that he feels he has no control over events
            The use of cut-scenes
                Printing large quantities of text in Inform 7
    High-suspense scenes (combat, physical danger, etc)
        Every turn rules that keep events rolling (see Heroine’s Mantle)
        Timed puzzle scenarios
        Impact of SAVE/UNDO on these scenes; some discussion on whether to disable (concluding: please don’t!)
        Providing lots of clues about what to do next so that the player doesn’t trip in these scenarios: if we want a scene that plays like an action scene or swash or something, the player should spend minimal time thinking about what to type next
    Trapped/set-length scenes (being locked in a cell, on a journey, etc)
        Pacing when the player is faced with a single problem to solve
            Focus, controlled frustration
            Opportunity to convey information, esp. if the player is unable to leave while he hears/sees an event (see Christminster, Delusions)
        Legitimate ways to trap the player
            Preventing actions during a scene — using “during”
            Writing sensible refusals while the scene progresses
        Maintaining interest while working on the single problem
            Designing puzzles with feedback for partial solutions
            Atmospheric techniques to maintain local color (see Atmosphere)
            Internal monologue, development of character reaction to situation (see also Viewpoint and Viewpoint character)
        Scenes with a minimum length where player cannot finish in less than N turns
    Exploratory scenes
        Leaving hints and guidance to move the player along (“Entrapment”)
        Exploration with a specific goal in mind, giving shape to a sequence
            Descriptions and reactions that reflect what player already knows
            Tracking what player has seen (perhaps introduce Epistemology extension, “Prague Job”, “Unexamined Life”, “Solitude”)
    Conversation scenes
        Switching control between the player and the non-player character
        Different functions of conversation
        (Probably would leave the particulars of this until later, since it’s complicated to implement and depends on your conversation model)
    Ending scenes with a hook
        Making the final action of a scene important (and something that feeds into the later parts of the game)

Pacing with Puzzles; Puzzle Structure

    Puzzle pacing vs. scene pacing; concepts of linearity and breadth
    Geographical puzzles
        Puzzle types supported by the existing library; standard puzzle design issues
            locks and keys (chapter 3; Locksmith)
            Characters guarding spaces
            Light (chapter 3; “descriptions of a dark room” in activities chapter)
        Designing a map with pacing in mind
    Research puzzles
        Making sure the player has learned background before moving on
        Tie back to “in-game sources of information” above, perhaps
    Structuring the plot around time blocks in which different things are possible (Anchorhead, Christminster)

Conflict and Motivation

    Setting goals for the viewpoint character
        Problems in forcing viewpoint character into one emotional state or another
        Leaving goals partly open, including some element of choice (not a required feature of a work of IF, but again worth thinking about)
        Refer back to exposition section for some ways to give a character background history
    Conflict with self/internal urges (see The Baron, Shrapnel, Rameses)
        Design issues to do with forcing player action / taking away some player control
        “try doing x” and other action-forcing gimmicks; cut-scenes
    Conflict with other characters (see Plundered Hearts, Shadows on the Mirror, Spider and Web, Elysium Enigma)
        Standard puzzles involving characters blocking player progress; missions for characters
        Withholding or offering information to characters; struggles over knowledge and information (anticipates stuff about modeling character knowledge, but we’ll get to that later)
        Modeling characters who are actively working against the player (When in Rome 2; in less simulationist terms, Gourmet)
        More on conversation later
    Conflict with the natural environment (see Hunter, in Darkness, I-0)
        Refer back to use of geography and setting
        Physical modeling; appropriate degrees of detail; the fine line between fun and frustration
        Ways to do timed puzzles; “MRE”; background on (deprecated) hunger, sleep, and disorientation/maze puzzles
        Possible more sophisticated ways to do struggle-against-nature
    Conflict with a social system or circumstance (see “Wishbringer”, Kaged, Square Circle)
        Repression as expressed by limitations on player action
        Use of strongly-reactive other characters who respond rapidly to deviant behaviors
        Use of frequent (undoable) death as a way to heighten stress; death and afterlife commands in Inform 7

Viewpoint and the Viewpoint Character

        Physical description
            Describing the viewpoint character in response to X ME
            Giving the viewpoint character appropriate clothes
                Creating inventory; controlling description within inventory
                Selecting what to model depending on the focus of the piece, since not every character needs a full set of clothes, but this is sometimes appropriate to have for some characters
                Giving viewpoint character appropriate tools and props
                communicating what the character is and can do from the outset
        Action refusal messages: what the character won’t do reveals who he is (see Rameses)
            Using extensions or rules to change the default library messages
        Room descriptions and character attitude (see Common Ground)
        Allowing the player to role-play and explore the character
            Custom actions reflecting character-appropriate behavior (see Tale of the Kissing Bandit, The Act of Misdirection)
                Creating simple new actions
                Cluing character-appropriate commands in the text
    Using internal monologue
        Modeling character’s attitude and discoveries
                Using relations and/or epistemology extension to track
                Variables representing mood or other state (see Sunset Over Savannah)
            Representation to the player
                Triggering internal thoughts automatically
                    Using every turn rules to interject thoughts independently of player action
                Triggering thoughts as a response to player actions
                    Writing special response/instead rules for special situations
                Allowing the player to explore mental state interactively
                    Implementing commands involving thought: THINK, REMEMBER; “Merlin” example
                    Writing commands that apply to things that aren’t visible; scope manipulation, representing abstracts
                Allowing the player to express mental state for the character
                    Implementing common gestures (even if they don’t much affect the surrounding world)
                        KICK, SMILE, FROWN, SHOUT, ETC.
                        Making sure these behave appropriately around other characters — sometimes these may be more trouble than they’re worth
    Changing the viewpoint character
        “change the player to…” stuff (see Being Andrew Plotkin; past raif threads on this)
        Changing room descriptions etc. accordingly (also covered above)
        Using Custom Library Messages for first/third person (see Fallacy of Dawn)

Other Characters

    Reactive characters who respond to/notice what the player does
        “in the presence of” and related rules; “Day for Fresh Sushi” example
        Physical interaction with characters
            Standard library actions (KISS, ATTACK, SHOW, GIVE)
            Combat and action sequences
                Some problems with randomized combat in IF
                (May refer back to action-scene stuff earlier)
            Sex sequences
    Characterization in absentia
        Physical artifacts, memories, etc., giving clues about characters
        Basic action stuff on conversation
            Choosing a conversation model for your work
            How to use ask/tell (chapter on Basic Actions; Conversation Rules extension)
            How to substitute in TALK TO
            How to use conversation menus; Simple Chat extension
            Advanced conversation model design
                Actions that apply to abstract/otherwise invisible objects; scoping
                Representing relationships between topics and statements
                Organizing data with relations and tables; sorting tables; nesting table references
        Writing good conversation output
            Giving player topics to follow up on, in ask/tell scenarios; “Chronic Hinting Syndrome”
            Presenting gesture and attitude of the character
            “Talking Head Avoidance Device”
                Involving the surroundings in the conversation
                Letting the character do minor actions while talking
        Using scenes to structure conversation
    Automated character behavior
        Creating actions for other characters to perform (“Advanced Actions”)
            Using abstract actions such as “dine” or “fidget” which the code can then resolve into more specific acts (“The Man of Steel”)
            Goal-seeking (but talk about how this is likely to affect the author’s control over the story; design appropriately)
                Chaining before rules to produce basic goal-seeking behavior (“IQ Test”, “Boston Cream”)
                More sophisticated multi-track goal-seeking using Reactive Agent Planner
                Concept of plot-oriented goal-seeking rather than goal-seeking by NPCs
        Tracking character knowledge
            Relations, again, as a way to keep track of things
            Modeling information that is more than two-way (e.g., knowledge held with degrees of certainty)
« Last Edit: February 24, 2023, 01:34:09 PM by gimymblert » Logged

Level 10

The archivest master, leader of all documents

View Profile
« Reply #491 on: February 24, 2023, 09:46:42 PM »

Dramaton: Toward a Testbed for Exploring Interactive Storytelling

Level 10

The archivest master, leader of all documents

View Profile
« Reply #492 on: February 26, 2023, 09:13:15 AM »

Tightening the World-Plot Interface: or, Why I Am Obsessed With Conversation Models

Dynamic narration, simulationist and narrativist approach, and yandere simulator (by me)
Simulationist is all about the chaos as story, every agents pursue actively its own agenda and story occasionally emerged from it, it generally need a high level overview of story to extract the good bit, which is why shadow of mordor had to have an interface dedicated to showing shifting alliance and balance of power. And why the best example pointed by simulation are spreadsheet game. The metric of quality for a simulation is the depth and fidelity of the simulation, everything else is irrelevant. That's great, but that's not the only quality to pursue.

Narrativist is about the presentation layer and authorship, which simulationisty game lack. The quality is known as "well formed story". As such Narrativist aren't trying to simulate word with competing agent, but simulate the story space of meaning, plotting, stakes and character functions working toward the "well formed story".

As such narrativist see simulation as herding cat and created the concept of the "drama manager" that instruct agent which role they play. Looking back, the simplest drama manager is probably a schedule table, character have goal based on the time of day rather than deciding from internal characteristic. The table can be anything such as an atmosphere level, the story progression state, etc ... agent implementing and selecting reactive behavior contextual to the goal specified in the table based on their traits (persona in my case).

Drama manager can be more complex and reason or react on the state of the world, to implement authored story beat, by adding, removing or shifting role of agents. Like an RTS ai but dealing with "well formed story" as a goal. Generally you will have a background simulation, that is stable because of the table, and the player perturbing the stability, and pushing the story world to new state, forming a new plausible story. That is the story emerge from and around the player's actions, which mean it doesn't need an extra spreadsheet to extract a story. It's the concept of elastic story, it's a akin to blended animation tree applied to story.

Also you are wrong when you say it hasn't been done before, but I'm not sure you want a history lesson on the various implementations, just gonna say you could have least study the list of social simulation you send me, Yandere simulator is an evolution (direct inspiration) from Tokimeki memorial, it also have elements from façade, versu, and Bethesda's radiant ai.

The reason I'm cloning it, is to find where I could tacked on other model (especially conversation) to make it more complete. The reason why I try to separate architecturally the generic behavior and the specific behavior, is to have a structure more like versu's (character file and genre file, story file is already there in the form of the global agent manager) but 3d.

The reason it worked is not because the dev is smarter than anyone else, but because the format of gameplay is ideal to make the case, it has unity of actions, unity of place and unity time. It all happen in a single place, the schools, where all characters stay the same and has low variability (students and teacher). That mean, that unlike RPG, you have high frequency interactions with the same sets and class of character, it has density.

RPG have lots of character class you only meet transiently, spread over a large world, with complex technical requirement like LOD that doesn't allow them to be simulated at all time, that mean efforts is spread horizontally, and (story/world) states aren't correctly conveyed. Smaller game like versu and overboard, have issues with high surface of input (the entire language), that cover a surface of actions too big to be adressed correctly, and have even more class of character with all character being unique.

Yandere simulator strike the nice balance between complexity and scope to surface a working structure we can expend from. And like I said the format it has did help a lot, it has generic character lift from anime tropes, which mean they are readily legible, you don't have to guess their personality and actions, making them perfect illustration for the implementations. It has a hierarchy of importances between agents, with regular school kids, teacher and rival, who have different roles in the story world, with regular students as background, being both obstacles and resources.

The actions is entirely contextual on the social space, every action is significant, crouching in the corridor makes you suspicious, being at a place at the wrong time have consequences, holding objects at different time mean different things, and the generic settings makes it legible, it makes sense to not run with a mop during lunchtime, and during clean time nobody will ask you why you hold plastic bag.

Even doing nothing have a consequences in the stakes of the story world, and the progression is handled by simply switching simulation state (like school in high alert, where camera is placed, students acting distrustful, happy go lucky photography club improvising themselves as detective, all base on a mix of generic behavior and specific behavior selected by the general agent manager). And how you act will be tested against you in the ending trial, so all actions taken will come back in a reverse detective sequence in which you try to prove your innocence, which integrate further the notion of "well formed story".

I learned a lot studying it, and I believe we can perfect the lesson learned. We can make more complex story and more complex character and not leave everything to dialog tree. That is a game in which you ACT, that is visual medium have "show don't tell", gamist will told you "do don't show", and this will open the case for "BE don't do".

It act as a working missing link between various implementation, and the thing that interest me, it's not a text heavy solution created around conversation, but a 3d world like regular games in which you communicate with direct actions, which allow to generalize the model and bridge the gap with common actions gameplay. Basically it allows to conceptualize how to port a model like Versu toward game like mass effects.

Level 10

The archivest master, leader of all documents

View Profile
« Reply #493 on: February 26, 2023, 02:21:03 PM »

Standard Patterns in Choice-Based Games

[email protected]
« Reply #494 on: February 26, 2023, 08:16:37 PM »

Dialogue trees aren't really an AI problem unless you make them an AI problem, which I have yet to see. I'm tired of people settling for bloated SaaS models like GPT and just generating things, rather than embedding state-of-the-art AI in their games. Meta's smallest usable chatbot is 730 MB for the model + however much bloat is needed for the Transformer framework. I know we can do better.
Level 10

The archivest master, leader of all documents

View Profile
« Reply #495 on: February 27, 2023, 01:37:45 PM »

Dialogue trees aren't really an AI problem unless you make them an AI problem, which I have yet to see. I'm tired of people settling for bloated SaaS models like GPT and just generating things, rather than embedding state-of-the-art AI in their games. Meta's smallest usable chatbot is 730 MB for the model + however much bloat is needed for the Transformer framework. I know we can do better.

Depend if old school dialog tree or rule based quips, which has precondition, metadata, and allow npc to planned through the conversation graph.

Survey of Storylets-based Design
Storylets: You Want Them
Mailbag: High-Agency Narrative Systems

 Welcome to the Storychoices wiki
StoryNexus creators’ pack: public beta edition
Montage, Narrative Deckbuilding and Other Effects in StoryNexus

« Last Edit: February 27, 2023, 03:44:45 PM by gimymblert » Logged

[email protected]
« Reply #496 on: February 27, 2023, 02:16:30 PM »

Those are examples of a human writer writing all the dialogue, which tends to railroad the story. I think there are better alternatives.
Level 10

The archivest master, leader of all documents

View Profile
« Reply #497 on: February 27, 2023, 03:16:43 PM »

 "Ceptre: A Language for Modeling Generative Interactive Systems" by Chris Martens

Level 10

The archivest master, leader of all documents

View Profile
« Reply #498 on: February 27, 2023, 03:19:28 PM »

Those are examples of a human writer writing all the dialogue, which tends to railroad the story. I think there are better alternatives.
A story is railroad by definition, else it's not a story. Anyway, every world is finite, I create stuff with chat gpt, and he goes directly for tired trope, so even generative ai that has read all the internet won't escape a limitation. The proble is that meaning IS repetition, so if you want something meaningful, you will limit yourself in any way, else you will have to accept noise as valid creation.

Level 10

The archivest master, leader of all documents

View Profile
« Reply #499 on: February 27, 2023, 05:51:47 PM »

Exercises in Generated Prose

Modeling conversation flow: types of NPC Initiative

Modeling conversation flow: subject changes

Modeling conversation flow: transitions in player speech

The beat issue

Modeling conversation flow: interrupting the player

Modeling conversation flow: NPC repeating information

Modeling conversation flow: NPC questions

Modeling conversation flow: actions in conversational context

Modeling conversation flow: multiple people conversing; some closing caveats

The conversational “script”

Modeling conversation flow: beginnings and endings using scenes

Conversational analysis studies

« Last Edit: February 28, 2023, 12:43:15 PM by gimymblert » Logged

Pages: 1 ... 23 24 [25] 26 27 ... 30
Jump to:  

Theme orange-lt created by panic