apoch / curvature

A full-featured editor for working with Utility-based AI

Curvature

Curvature is a full-featured Utility-based AI editor and sandbox tool.

The project provides a complete playground for creating, editing, and testing decision-making AI. Under the hood, Curvature uses Utility Theory to model the appeal of various behaviors. Specifically, Curvature is based on the Infinite Axis Utility System by Dave Mark. Curvature builds on the IAUS approach and includes enhancements and refinements developed during work on Guild Wars 2: Heart of Thorns as well as Guild Wars 2: Path of Fire.

Curvature is fully data-driven and supports the creation of a complete AI pipeline, from the core knowledge-base accessed by agents, to the specific considerations that drive the scoring of individual behaviors. The result is an end-to-end solution for modeling and testing AI, including a simple world representation that allows designers or AI programmers to place agents in a virtual space and see how they would choose to behave.

For complete project documentation, please see the Curvature wiki.

Screenshots

Project Status

Curvature is currently in an open beta testing phase. Check the project releases page for the latest version of the tool suite.

Existing features of Curvature:

• Design a knowledge-base to contain your world representation and data
• Select knowledge-base entries to assemble a list of inputs that can drive decision-making
• Pair an input with a decision using a consideration, which can control the relevance of the input via response curves
• Collect considerations to make up a behavior which represents a discrete choice available to an AI agent
• Group behaviors into behavior sets for easy categorization and bundling of related activities
• Choose which behavior sets will be enacted for a given archetype of AI agent
• Place AI agents in a sandbox world and watch them make decisions using scenarios
• Response curves and considerations can be constructed with a simple wizard-style helper
• Running scenarios will log the decisions made by each agent at each "tick" so you can review how things came about

More fine-grained information on project plans can be found in the issue tracker.

Overviews of releases and other high-level planning can be found in the project tracker.

curvature's Issues

Bug: selection of edited object is lost after renaming it

Have rename events maintain the selected item in each appropriate listview.

Bug: curve presets don't apply correctly

When using the Curve Preset dialog, the selected preset is not successfully applied to the project even if OK is clicked.

Feature: dirty flag

Implement project "dirty" flag such that edits flag a project as dirty and saving clears the flag.

Polish: fill whitespace when no behavior set selected

Currently the UI shows a blank area when no behavior set object is selected (including but not limited to when no object exists). Come up with a better solution for this that makes effective use of space.

Bug: scenarios do not set the project "dirty" flag

Editing content in a Scenario does not mark the project as having unsaved changes.

Feature: properties on locations

Scenario locations need properties for things like tagging.

Bug: manual simulation advancement runs more than 100ms of sim time

Several seconds seem to elapse for a couple of clicks on Advance 100ms. Double check the plumbing.

Bug: project statistics do not refresh in realtime

Creating a bunch of content does not cause the Project tab stats to refresh directly; other events can force a refresh (such as renaming the project or saving/reloading) but it'd be preferable if the numbers were always accurate. Just for fun.

Feature: delete considerations

Add UI functionality for deleting considerations from a behavior.

Polish: fill whitespace when no archetype selected

Currently the UI shows a blank area when no archetype object is selected (including but not limited to when no object exists). Come up with a better solution for this that makes effective use of space.

Bug: sliders with a range of [0,1] don't work

Set up an input axis with range of [0,1] and clamp its consideration range to 1. The input slider on the behavior scoring tab (among other places) will be busted.

Feature: improved rendering in scenarios

Implement a few features to make the scenario rendering nicer:

• Configurable colors for each agent/location
• Implement and display facing-direction for each agent
• Add simple animations for actions/intentions

Polish: keyboard shortcuts

Implement keyboard shortcuts for menu items and common navigation actions/etc.

Bug: inputs do not reflect their properties correctly

Currently when editing an input axis there is UI for selecting exactly what inputs will be used (and where they will come from). Unfortunately this UI does nothing.

Bug: Incorrect attempt to union inputs

If two inputs are provided that read the same KB record but use a different origin (me vs. target) the Union function will still be called and (rightfully) barf out an exception.

Bug: clicking Cancel in the consideration wizard does not roll back changes made

Repro:

• Create a consideration on a behavior
• Input some values in the wizard
• Click the consideration in the list of considerations for the behavior
• Click the Wizard button to the right of the Input dropdown
• Scramble some values
• Click Cancel instead of OK
• De-select and re-select the behavior

The scrambled values will be reflected in the consideration editor even though the dialog was canceled. This should instead roll back the dialog's changes to the state the consideration was in prior to opening the wizard.

Polish: add visuals to "new object" buttons

Add some kind of iconography to the buttons that create objects.

Polish: knowledge-base widget layout

Consider moving all controls into a single horizontal row instead of the current two-row split.

Polish: consistent terminology/naming pass

Do a pass on all UI and make sure terms are consistently used.

Polish: fill whitespace when no behaviors selected

Currently the UI shows a blank area when no behavior object is selected (including but not limited to when no object exists). Come up with a better solution for this that makes effective use of space.

Feature: delete scenario agents

Scenario agents need a Delete button.

Feature: data persistence

Implement loading/saving of projects. Probably worth waiting on this until most of the functionality is in place to avoid versioning nightmares.

Feature: add parameterization of inputs

Currently all inputs are set up to assume a fixed range. Extend support for parameterizing ranges such that inputs can have a custom range on each consideration.

As part of this, make sure that the higher-level experimentation modes merge the ranges correctly, i.e. all possible values for an input should be settable from the experimentation UI without needing a slider per range.

Feature: archetype editing

Needs design, but add a UI layer for setting up archetypes of characters.

Feature: record properties of agents/locations in scenario log view

The logs of decisions in scenario mode do not record the (raw) input values of various properties used to make decisions; only the normalized input scores. See if this is viable to add.

Feature: random seeds for scenarios

The Random Seed dropdown UI should be completed to support actual selection of a random number seed for a simulation. Tracking a history of selected seeds would be cool.

Feature: randomize agent start locations

Add a checkbox in Scenarios so that agents can be given a random (based on a user-specified seed!) start location. Probably needs a "suggested boundary" concept or something similar to help keep things from going crazy.

Polish: code cleanup pass

Do a thorough code review and clean up all the annoying stuff.

Bug: clicking the Edit Name icon twice causes double edit fields to appear

• Pick a thing to rename, such as an Input Axis
• Click the rename icon (currently a blue pencil)
• Click the icon again

A pair of input fields will appear and the rename functionality will start behaving badly.

Polish: fill whitespace when no scenario selected

Currently the UI shows a blank area when no scenario object is selected (including but not limited to when no object exists). Come up with a better solution for this that makes effective use of space.

Bug: UI buttons on Behaviors, Behavior sets, Archetypes and Scenarios type almost not accessible. (Win10)

As demonstrated in the image below:

The button to create Behavior/Archetype/Scenario is almost not accessible.
Also, the Considerations UI inside Behaviors Tab (Edit Consideration) is not visible.
Something similar happens in the Behavior set tab (right part)

Changing the size of the window does not solve the problem.

Feature: renaming objects

Add context menu item and other applicable UI to enable renaming of content objects.

Bug: consideration wizard Parameterization tab text is wrong for enumerations

When a consideration uses an input axis that has an enumeration constraint, the Consideration Wizard dialog talks about scaling the input value from 0 to 1 by dividing by the parameter. This prose is correct for fixed-range params but wrong for enumerations. Detect if the parameter being chosen is an enumeration and dynamically change the prose as necessary.

Bug: renaming objects does not trigger UI refresh

Any nameable object should trigger UI updates when it is renamed.

Polish: finish/revise prose in Curve Wizard

The consideration wizard text needs some attention, especially the last tab on naming.

Feature: compensation formula

Implement the many-axis compensation formula into the scoring system, so that decisions with many inputs are not artificially penalized.

Feature: enumeration-typed parameters

Finish design and implementation/UI for enumerated parameters.

Feature: sliding response curves

Add offset adjustment buttons and/or click-and-drag support to the response curve editor.

Feature: curve wizard

Introduce a guided step-by-step process for creating new response curves. Should be a complete solution for adding curves with various common shapes. Probably encompasses/supplants preset curve UI.

Feature: edit/delete scenario locations

Scenario locations can currently be added but not modified or removed.

Feature: scenarios

Build a UI for placing characters in a simple 2D world and having them "think" through their respective behavior sets.

Feature: Project overview page

Brainstorm content and implement something cool for Project Overviews.

Polish: Scenario Logs numeric precision

Adjust the output of numbers in Scenario Logs to 3 decimal places.

Feature: behavior sets

Rearrange the data hierarchy (and UI) so that behaviors are a soup and behavior sets are assembled piecemeal from the list of existing behaviors. Consider adding a "behavior category" ply to the hierarchy, so that the behavior soup is not unwieldy.

Feature: finish deletion support

Complete the implementation of object deletion for:

• Behavior sets
• Archetypes
• Scenarios

Task: catalog remaining wiki documentation and write it

A number of hint links in the UI point to non-existent wiki documentation. There are also pages on the existing wiki that don't link to anything because the thing they should link to has not yet been written.

Do some writing!

Polish: tab order pass

Do a complete pass over all UI elements for tab ordering.

Polish: consistent variable/member names

Some concepts are named inconsistently, e.g. "EditProject" vs "EditedProject" vs "EditingProject" etc.

Feature: scenario inspectors

Add inspection capabilities so that agent thought processes can be monitored and debugged during the playback of a scenario.

Bug: many edits do not propagate fully through the UI until something is re-clicked

One example: create a behavior and add a consideration to it. Rename the consideration. The list of considerations does not refresh until the behavior is re-clicked.