Reinforcement Learning Thermostat Controller Comparison

The team compared three trained reinforcement-learning policies—Q-learning, SARSA, and DQN—with Rule-Based and Proportional fixed controllers in one simplified single-zone room model. Off and High were retained as non-practical reference bounds rather than candidate thermostat designs.

The comparison records mean reward, absolute temperature error, normalized control effort, and action switching. Those measures answer different questions: tracking a setpoint closely does not by itself establish quiet control action, low real-world energy use, or deployment readiness.

02 — Room model and reward

The shared plant, action space, and reward define both the fairness and the limits of the controller comparison.

At each simulated step, room temperature changes through the selected normalized HVAC action, heat exchange with a fixed 22°C environment, and Gaussian disturbance noise. Every controller chooses from the same five actions, spanning strong cooling to strong heating: -1, -0.5, 0, 0.5, and 1.

The reward penalizes absolute deviation from the 25°C target and adds 0.1 times action magnitude. It contains no penalty for changing actions, which matters when the final switching counts are interpreted later.

Documented simulated plant and reward

The first relation is the common single-zone temperature update. The second is the project-specific reward: temperature error and normalized action magnitude are penalized, while action switching is not.

: Simulated room temperature at the current step.
: One of five normalized cooling/heating actions.
: Fixed simulated ambient temperature, 22°C.
: Simulated target temperature, 25°C.
: Gaussian disturbance with mean 0 and standard deviation 0.2.

The final standardized model uses α = 0.6 and β = 0.015. The source assigns no physical duration to a step and no physical units to these coefficients or normalized actions.

Final standardized simulation parameters
Quantity	Documented value	Interpretation
Target / ambient	25°C / 22°C	Fixed simulated temperatures
Response α / loss β	0.6 / 0.015	Source-defined coefficients; no physical units assigned
Noise	Gaussian μ = 0, σ = 0.2	Simulated disturbance, not a calibrated sensor model
Actions	[-1, -0.5, 0, 0.5, 1]	Normalized cooling-to-heating levels
Temperature range	10–40°C	Initial-state and tabular-discretization range
Episode horizon	200 steps	No physical time per step is documented

The modeled action magnitude is unitless normalized control effort—not kW, kWh, electrical consumption, or operating cost.

03 — Five controllers, two reference bounds

Three learned policies and two fixed traditional baselines form the main comparison; Off and High provide context at the extremes.

Q-learning and SARSA learn tabular action values from current temperature discretized into 50 bins across the documented 10–40°C range. DQN instead maps one normalized continuous-temperature input to five action-value estimates; the source does not preserve its layers, activations, optimizer, replay settings, target-network schedule, batch size, or learning rate.

The fixed controllers require no training. Rule-Based applies full heating below 24.5°C and otherwise selects no action. Proportional maps five documented temperature-error bands to the same five actions. Off always selects zero and High always selects maximum heating, so they remain contextual bounds rather than practical controllers.

Controller methods
Policy	Type	State / rule	Training status
Q-learning	Tabular off-policy RL	Current temperature · 50 bins over 10–40°C	Trained
SARSA	Tabular on-policy RL	Current temperature · 50 bins over 10–40°C	Trained
DQN	Neural action-value RL	One normalized continuous-temperature input · five outputs	Trained
Rule-Based	Traditional fixed logic	+1 below 24.5°C; otherwise 0	Not trained
Proportional	Traditional fixed logic	Five temperature-error bands mapped to the common actions	Not trained
Off	Reference bound	Always 0	Not trained
High	Reference bound	Always +1	Not trained

Off and High are reference conditions, not practical thermostat controllers. The available project record does not specify enough DQN training detail to reproduce that controller fully.

04 — Train first, compare on matched episodes

Policy learning and controller evaluation are separate stages with different counts and evidentiary meaning.

Q-learning, SARSA, and DQN each trained for a fixed 2,000 episodes of 200 steps in the common environment. The documented shared settings include discount γ = 0.95, epsilon-greedy exploration from 1.0 with decay 0.995 and minimum 0.05, and a 0.1 learning rate for the two tabular methods. The source does not document a DQN learning rate or formal convergence criterion.

After training, all seven policies were evaluated over 100 new simulated episodes of 200 steps. Evaluation seed 12345 generated matched initial temperatures and noise sequences for every controller, making the within-simulation comparison consistent without establishing a held-out environment, independent training repeats, or out-of-sample generalization.

Training and evaluation are distinct stages
Stage	Policies	Documented protocol	Limitations
Training	Q-learning, SARSA, DQN	2,000 episodes × 200 steps · base seed 42	One fixed schedule; no formal convergence or multi-seed study
Evaluation	All five controllers + Off / High	100 matched episodes × 200 steps · seed 12345	Descriptive means in the same model family; not a held-out test

The 100 evaluation episodes are not 100 independently trained models. Per-episode standard deviations and confidence intervals were not preserved.

05 — Accuracy and control effort are different questions

The final means expose a tradeoff that a single winner label would hide.

Five-controller evaluation means

Bar chart comparing mean absolute temperature error for Q-learning, SARSA, DQN, rule-based, and proportional controllers over 100 simulated evaluation episodes. — Average absolute temperature error · °CMean absolute deviation from the 25°C target across 100 matched evaluation episodes. Exact values appear in the table below.

Bar chart comparing mean normalized action magnitude for Q-learning, SARSA, DQN, rule-based, and proportional controllers over 100 simulated evaluation episodes. — Average normalized control effort · unitlessMean normalized control effort, calculated as action magnitude |u_t|, across 100 matched episodes. Despite the source figure title, this is not physical energy use.

DQN recorded the lowest mean temperature error among the five main controllers, while both traditional baselines used less normalized control effort and switched less often.

Final standardized seven-policy evaluation means
Policy	Reward	Error (°C)	Normalized effort	Switches / 200 steps
DQN	-86.05	0.408	0.227	97.13
Q-learning	-97.98	0.436	0.537	126.00
SARSA	-98.67	0.440	0.537	125.87
Proportional	-107.31	0.523	0.137	51.81
Rule-Based	-194.97	0.965	0.095	24.33
Off (reference)	-704.47	3.522	0.000	0.00
High (reference)	-5093.32	25.367	1.000	0.00

Values are computed descriptive means across 100 matched simulated episodes. Reward is specific to the documented objective; normalized effort is unitless; no variance or significance result is available.

DQN under this reward

-86.05 reward · 0.408°C mean error

DQN produced the highest mean reward and lowest mean absolute temperature error in the final standardized evaluation.

This result applies to the simulation and reward used here. It does not establish statistical significance or general superiority.

Strongest reported traditional baseline

Proportional · 0.523°C error · 0.137 effort

Proportional control accepted higher mean error than DQN while using less normalized control effort and changing action less often.

Normalized effort is a simulated action-magnitude metric, not electrical consumption.

No confidence intervals, statistical-significance test, robustness study, or real-building measurement is documented.

Q-learning and SARSA tracked the target closely in mean error, but each used 0.537 normalized effort and averaged about 126 action changes per episode. DQN reduced both values relative to the tabular methods, yet still changed action substantially more often than either traditional baseline.

The Off and High rows make the reference bounds explicit without compressing the five-controller charts. Zero action does not hold the target in this model, while constant maximum heating performs poorly under the same reward.

06 — Action switching and next steps

Frequent simulated action changes reveal an objective gap that must be addressed before physical-device claims are possible.

Q-learning and SARSA averaged 126.00 and 125.87 action changes per 200-step episode; DQN averaged 97.13. Proportional and Rule-Based control averaged 51.81 and 24.33. These are simulated action-count observations, not measured hardware cycles, wear, instability, or maintenance cost.

Because the documented reward penalizes temperature error and action magnitude but not action changes, the switching result identifies a missing objective. Adding a switching penalty, weather-driven inputs, multi-zone physics, hardware constraints, and physical testing remain future work rather than completed iterations.

Learned controllers

97.13–126.00 switches / 200 steps

All three learned policies changed action more often than both traditional baselines in the final evaluation means.

The model defines no physical time per step, so these values cannot be converted to equipment cycles per hour.

Traditional baselines

24.33–51.81 switches / 200 steps

Rule-Based and Proportional control changed action less often while accepting higher mean temperature error than DQN.

The comparison does not establish hardware friendliness, safety, or physical stability.

I focused on validation and documentation, organized the figures and metrics, and wrote the validation and final comparison sections. The controller implementations, simulation, training code and baselines were developed across the team.

01 — A controlled thermostat comparison