SirStone
aa4bc77068
Two-hidden-layer MLP actor-critic (42→64→64→5/1) with stochastic actorForward, logStd floor at -3, and BotAction mapper wired into the run() loop. Assert-based test suite covers shapes, finiteness, logStd collapse, and all action range bounds. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
34 lines
1.2 KiB
Nim
34 lines
1.2 KiB
Nim
## actions.nim — map raw network output to Tank Royale bot commands.
|
|
|
|
import arraymancer
|
|
import std/math
|
|
|
|
type
|
|
BotActions* = object
|
|
targetSpeed*: float32
|
|
turnRate*: float32
|
|
gunTurnRate*: float32
|
|
shouldFire*: bool
|
|
firePower*: float32
|
|
|
|
proc mapActions*(rawActions: Tensor[float32], currentSpeed: float32, gunHeat: float32): BotActions =
|
|
## rawActions: [5] tensor from actorForward.
|
|
## Ranges:
|
|
## targetSpeed ∈ [-8, 8]
|
|
## turnRate ∈ [-(10 - 0.75|v|), (10 - 0.75|v|)]
|
|
## gunTurnRate ∈ [-20, 20]
|
|
## firePower ∈ [0.1, 3.0]
|
|
let r0 = tanh(rawActions[0].float64).float32
|
|
let r1 = tanh(rawActions[1].float64).float32
|
|
let r2 = tanh(rawActions[2].float64).float32
|
|
let r3 = tanh(rawActions[3].float64).float32
|
|
let r4 = rawActions[4]
|
|
|
|
result.targetSpeed = r0 * 8.0'f32
|
|
result.turnRate = r1 * (10.0'f32 - 0.75'f32 * abs(currentSpeed))
|
|
result.gunTurnRate = r2 * 20.0'f32
|
|
let fireDecision = r3
|
|
result.shouldFire = fireDecision >= 0.0'f32 and gunHeat <= 0.0'f32
|
|
# sigmoid(r4) * 2.9 + 0.1 → [0.1, 3.0]
|
|
result.firePower = (1.0'f32 / (1.0'f32 + exp(-r4.float64).float32)) * 2.9'f32 + 0.1'f32
|