tommy/tbd-station-14
2
1
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Geometric mean for AI considerations (#1812)

Browse Source 
* Geometric mean for AI cons

More forgiving for having more considerations for an action.

* You had 1 job

Co-authored-by: Metal Gear Sloth <metalgearsloth@gmail.com>
This commit is contained in:
metalgearsloth
2020-08-20 22:16:12 +10:00
committed by GitHub
parent 599b316212
commit d8a78f2de6
2 changed files with 16 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
Content.Server/AI/Utility/Actions/UtilityAction.cs
View File

@@ -112,19 +112,14 @@ namespace Content.Server.AI.Utility.Actions
UpdateBlackboard(context);
var considerations = GetConsiderations(context);
DebugTools.Assert(considerations.Count > 0);
// I used the IAUS video although I did have some confusion on how to structure it overall
// as some of the slides seemed contradictory
// Ideally we should early-out each action as cheaply as possible if it's not valid
// Overall structure is based on Building a better centaur
// Ideally we should early-out each action as cheaply as possible if it's not valid, thus
// the finalScore can only go down over time.
// We also need some way to tell if the action isn't going to
// have a better score than the current action (if applicable) and early-out that way as well.
// 23:00 Building a better centaur
var finalScore = 1.0f;
var minThreshold = min / Bonus;
context.GetState<ConsiderationState>().SetValue(considerations.Count);
// See 10:09 for this and the adjustments
foreach (var consideration in considerations)
{

15
Content.Server/AI/Utility/Considerations/Consideration.cs
View File

@@ -13,10 +13,19 @@ namespace Content.Server.AI.Utility.Considerations
private float GetAdjustedScore(Blackboard context)
{
var score = GetScore(context);
/*
* Now using the geometric mean
* for n scores you take the n-th root of the scores multiplied
* e.g. a, b, c scores you take Math.Pow(a * b * c, 1/3)
* To get the ACTUAL geometric mean at any one stage you'd need to divide by the running consideration count
* however, the downside to this is it will fluctuate up and down over time.
* For our purposes if we go below the minimum threshold we want to cut it off, thus we take a
* "running geometric mean" which can only ever go down (and by the final value will equal the actual geometric mean).
*/
// Previously we used a makeupvalue method although the geometric mean is less punishing for more considerations
var considerationsCount = context.GetState<ConsiderationState>().GetValue();
var modificationFactor = 1.0f - 1.0f / considerationsCount;
var makeUpValue = (1.0f - score) * modificationFactor;
var adjustedScore = score + makeUpValue * score;
var adjustedScore = MathF.Pow(score, 1 / (float) considerationsCount);
return FloatMath.Clamp(adjustedScore, 0.0f, 1.0f);
}