feat(gamecli): implement pure Reward.Compute with unit tests
This commit is contained in:
56
GameCli.Tests/RewardTests.cs
Normal file
56
GameCli.Tests/RewardTests.cs
Normal file
@@ -0,0 +1,56 @@
|
|||||||
|
using GameCli;
|
||||||
|
using Xunit;
|
||||||
|
|
||||||
|
namespace GameCli.Tests;
|
||||||
|
|
||||||
|
public class RewardTests
|
||||||
|
{
|
||||||
|
[Fact]
|
||||||
|
public void AtTarget_ZeroVel_ZeroAngle_Noop_HasZeroReward()
|
||||||
|
{
|
||||||
|
var ship = new ShipState(0.5f, 0.5f, 0f, 0f, 0f, 0f);
|
||||||
|
float r = Reward.Compute(ship, target: (0.5f, 0.5f), action: 0);
|
||||||
|
Assert.Equal(0f, r, precision: 5);
|
||||||
|
}
|
||||||
|
|
||||||
|
[Fact]
|
||||||
|
public void FarFromTarget_HasLargerNegativeThanNear()
|
||||||
|
{
|
||||||
|
var near = new ShipState(0.5f, 0.5f, 0f, 0f, 0f, 0f);
|
||||||
|
var far = new ShipState(0.9f, 0.9f, 0f, 0f, 0f, 0f);
|
||||||
|
float rNear = Reward.Compute(near, target: (0.5f, 0.5f), action: 0);
|
||||||
|
float rFar = Reward.Compute(far, target: (0.5f, 0.5f), action: 0);
|
||||||
|
Assert.True(rFar < rNear, $"far reward {rFar} should be < near reward {rNear}");
|
||||||
|
}
|
||||||
|
|
||||||
|
[Fact]
|
||||||
|
public void HighVelocity_PenalizesReward()
|
||||||
|
{
|
||||||
|
var stopped = new ShipState(0.5f, 0.5f, 0f, 0f, 0f, 0f);
|
||||||
|
var moving = new ShipState(0.5f, 0.5f, 1f, 1f, 0f, 0f);
|
||||||
|
float rStop = Reward.Compute(stopped, target: (0.5f, 0.5f), action: 0);
|
||||||
|
float rMove = Reward.Compute(moving, target: (0.5f, 0.5f), action: 0);
|
||||||
|
Assert.True(rMove < rStop, $"moving reward {rMove} should be < stopped reward {rStop}");
|
||||||
|
}
|
||||||
|
|
||||||
|
[Fact]
|
||||||
|
public void Tilted_PenalizesReward()
|
||||||
|
{
|
||||||
|
var upright = new ShipState(0.5f, 0.5f, 0f, 0f, 0f, 0f);
|
||||||
|
var tilted = new ShipState(0.5f, 0.5f, 0f, 0f, 0.8f, 0f);
|
||||||
|
float rUp = Reward.Compute(upright, target: (0.5f, 0.5f), action: 0);
|
||||||
|
float rTi = Reward.Compute(tilted, target: (0.5f, 0.5f), action: 0);
|
||||||
|
Assert.True(rTi < rUp, $"tilted reward {rTi} should be < upright reward {rUp}");
|
||||||
|
}
|
||||||
|
|
||||||
|
[Fact]
|
||||||
|
public void FiringEngine_PenalizesReward()
|
||||||
|
{
|
||||||
|
var ship = new ShipState(0.5f, 0.5f, 0f, 0f, 0f, 0f);
|
||||||
|
float rNoop = Reward.Compute(ship, target: (0.5f, 0.5f), action: 0);
|
||||||
|
float rMain = Reward.Compute(ship, target: (0.5f, 0.5f), action: 2);
|
||||||
|
float rLeft = Reward.Compute(ship, target: (0.5f, 0.5f), action: 1);
|
||||||
|
Assert.True(rMain < rNoop, "firing main should be worse than noop");
|
||||||
|
Assert.True(rLeft < rNoop, "firing left should be worse than noop");
|
||||||
|
}
|
||||||
|
}
|
||||||
29
GameCli/Reward.cs
Normal file
29
GameCli/Reward.cs
Normal file
@@ -0,0 +1,29 @@
|
|||||||
|
namespace GameCli;
|
||||||
|
|
||||||
|
/// <summary>
|
||||||
|
/// Reward = -distance_to_target
|
||||||
|
/// - λ_v · speed
|
||||||
|
/// - λ_θ · |angle|
|
||||||
|
/// - λ_fuel · engine_on
|
||||||
|
/// Coefficients are public so training scripts can log/inspect them.
|
||||||
|
/// </summary>
|
||||||
|
public static class Reward
|
||||||
|
{
|
||||||
|
public const float LambdaVelocity = 0.10f;
|
||||||
|
public const float LambdaAngle = 0.10f;
|
||||||
|
public const float LambdaFuel = 0.03f;
|
||||||
|
|
||||||
|
public static float Compute(ShipState s, (float X, float Y) target, int action)
|
||||||
|
{
|
||||||
|
float dx = target.X - s.X;
|
||||||
|
float dy = target.Y - s.Y;
|
||||||
|
float distance = MathF.Sqrt(dx * dx + dy * dy);
|
||||||
|
float speed = MathF.Sqrt(s.VX * s.VX + s.VY * s.VY);
|
||||||
|
float engineOn = action == Physics.ActionNoop ? 0f : 1f;
|
||||||
|
|
||||||
|
return -distance
|
||||||
|
- LambdaVelocity * speed
|
||||||
|
- LambdaAngle * MathF.Abs(s.Angle)
|
||||||
|
- LambdaFuel * engineOn;
|
||||||
|
}
|
||||||
|
}
|
||||||
Reference in New Issue
Block a user