AI Trainer — F³ Racing

⚡ LIVE — PPO

SPEED 1×

LEARN RATE 3.0e-4

ENTROPY 0.0030

EPISODE LEN 60s

RANDOM SPAWN

↺ RESTART REQUIRED

AGENTS 8

🧬 PPO MODS (LIVE)

MIRROR AUGMENT KL EARLY STOP NEURON REPAIR

FAILURE RATE 0%

🏅 BEST NETWORK (AVG OF ALL AGENTS)

no snapshot yet

⚡ LIVE — REWARDS

PROGRESS /m 0.20

LAP BONUS 20

GRAVEL /s 1.0

WALL /s 2.0

CRASH-OUT 10

⚙ CONFIGURE TRAINING

SETTINGS ARE LOCKED ONCE TRAINING STARTS — RESTART REQUIRED TO CHANGE ARCHITECTURE

🧠 NETWORK ARCHITECTURE

HIDDEN LAYERS

more layers = learns complex patterns, trains slower

NEURONS PER LAYER

more neurons = more capacity, heavier updates

RECURRENT (GRU)
GRU memory cell trained by BPTT — replaces the memory-as-action register. Runs on CPU/WASM (no GPU).

…

🏎️ ENVIRONMENT

CAR

PARALLEL AGENTS 8 agents

more agents = more diverse experience per update (rounded to a multiple of the group size)

EPISODE LENGTH 60s

max seconds per run before reset

RANDOM SPAWN
start each run at random track position

ALL TRACKS (PARALLEL)
cars train across every map at once, rotating tracks each run — generalises the policy instead of overfitting one map. Use the VIEW button to watch a different track.

📊 TRAINING

AGENT GROUPS

groups spawn together at the same spot; each agent's advantage is its return vs the group average (GRPO-style — replaces the critic). Raise AGENTS so several groups run in parallel.

COMPUTE BACKEND

AUTO = multi-core WASM · GPU = TensorFlow.js on WebGL (stable everywhere, no WebGPU) · JS = plain fallback

GRADIENT THREADS 4 threads

CPU worker count for WASM/JS backends — more than your core count wastes memory

STEPS PER UPDATE (HORIZON)

more steps = better gradient estimate, slower update

LEARNING RATE 3.0e-4

how fast the network learns — too high = unstable

ENTROPY (EXPLORATION) 0.0030

higher = more random actions early on, helps explore