Q-learning hyperparameters

Learning rate α: 0.15 Discount factor γ: 0.90 Epsilon decay (per episode): 0.985 Training speed: 4 episodes/frame

Training control

Live playback

Playback speed: 1.0×

Training stats

Episodes trained0

Current ε (exploration)1.000

Last episode reward—

Best episode reward—

Avg reward (last 50 eps)—

Thermostat baseline avgcomputing…

Agent vs baseline—

Live greenhouse status

ModeIdle

Sim time—

Outside temp—

Greenhouse temp—

Humidity—

Heateroff

Ventclosed

Step reward—

Episode reward so far—

The agent's state is a discretised (temperature bin × humidity bin × time-of-day bin) and its actions are {heater, vent} on/off combinations. Every training episode genuinely updates Q(s,a) via the Bellman update Q(s,a) ← Q(s,a) + α[r + γ·max Q(s′,a′) − Q(s,a)]. The chart below is the real, unscripted episode-reward learning curve.