Demo for VAE Optimization

Pre-request

-Python 3.9 -Numpy -Pytorch 1.9.0 -jenkins -jenkinsapi -tqdm

Train Autoencoder and Approximate Reward Predictor (Only Driven Parameters) (New version no longer support)

User could mark the second command in run_train.bat and only run the first one. We provide some simulated data in dataset/data.csv. Trained weight of autoencoder and reward network are store in store/after_reward.pth and store/ae.pth. In the demo, reward is set as "max_hover_time_s".

Train Autoencoder for both Driven Parameters and Component Instance

Run the second command in run_train.bat. As no prior knowledge to know what makes a valid composition of parameters, data to train autoencoder are randomly sampled.

Train Reinforcement Learning for both Driven Parameters and Component Instance

Run the third command in run_train.bat. As no prior knowledge to know what makes a valid composition of parameters, data to train RL are randomly sampled. Information about whether different designs are valid are stored in file full_record.csv, which could be used to get rid of some simulation.

Bayesian Optimization to Optimize Only Driven Parameters

The first command in "run_optim.bat" optimize driven parameters with autoencoder to reduce dimension of search space and approximate reward network to predict reward.

Bayesian Optimization to Optimize both Driven Parameters and Component Instance

The second command in "run_optim.bat" optimize driven parameters and component instance. Similar as above, autoencoder is used to encode two kinds of parameters into continuous latent space. While jenkins is leverage for simulation.

Random Sample Test

Run the third command in "run_optim.bat".

Demo Results

Search space of component instance in Demo.

Component	Candidate Instance
Battery	TurnigyGraphene1000mAh2S75C, TurnigyGraphene1000mAh4S75C, TurnigyGraphene1300mAh4S75C, TurnigyGraphene1200mAh6S75C, TurnigyGraphene1400mAh3S75C, TurnigyGraphene1500mAh3S75C, TurnigyGraphene1600mAh4S75C, TurnigyGraphene2200mAh3S75C
ESC	ESC_debugging, kde_direct_KDEXF_UAS20LV, kde_direct_KDEXF_UAS55, t_motor_AIR_20A, t_motor_ALPHA_40A, t_motor_AT_40A,t_motor_AT_115A
Propeller	apc_propellers_6x4EP, apc_propellers_6x6EP, apc_propellers_7x4EP, apc_propellers_7x5EP, apc_propellers_8x8EP
Motor	t_motor_AS2312KV1400, t_motor_AT2321KV1250, t_motor_MN22041400KV, t_motor_MN4010KV475, t_motor_MT13063100KV

Best Parameters

Component Instance:

Component	Best Instance
Battery	TurnigyGraphene2200mAh3S75C
ESC	t_motor_AIR_20A
Propeller	apc_propellers_6x4EP
Motor	t_motor_AT2321KV1250

Driven Parameters:

Parameter	Value
ArmLength	200.047
SupportLength	121.610
BatteryOffset_X	-0.087
BatteryOffset_z	-0.656