intelligent-control-lab/agen

Adaptable generative prediction using recursive least square algorithm

Details for running AGen

Setup Environment

• install dependancy and data preprocessing see full_preprocessing_demon.md please modify dataset file name in validate_utils.py & adaption.py

• Install AGen clone files in this reporsitory directly to YOURPATH/ngsim_env/scripts/imitation

• Pretrained model and hyperparameters copy .npz files in ./pretrained/ to YOURPATH/ngsim_env/data/experiments/multiagent_curr/imitate/log/

Run Code

# Train and run a single agent adaptive algorithm
python adaption.py --n_proc 1 --exp_dir ../../data/experiments/multiagent_curr/ --params_filename itr_200.npz --use_multiagent True --n_envs 22 --adapt_steps 1(or2) 
# Train and run a single/multi adaptive algorithm
python adaption.py --n_proc 1 --exp_dir ../../data/experiments/multiagent_curr/ --params_filename itr_200.npz --use_multiagent False --n_envs 1 --adapt_steps 1(or2)

Supporting files

• theta.npy extracted top layer for pretrained RNN, use as initialization

• check_convergence.py check whether pretrained GAN is valid

Data Generation

output .npz file will be in YOURPATH/ngsim_env/data/experiments/multiagent_curr/imitate/

Data Analysis

a reference analysis example see data_analysis.ipynb

agen's Issues

Error Estimation in RLS adaptation code seems to be incorrect

Hello,
I'm a little confused about the error-estimation in RLS adaptation code.
I think Error Estimates in RLS adaptation is incorrect, but it may be that I misunderstood it.

39~52 lines of rls.py estimated the error after updating the parameter theta. Which is Posteriori Error Estimation:
pos_Error(t) = y(t) - phi(t)*theta(t+1)
But in the many real scenario, we need to report Priori Error Estimation:
pri_Error(t) = y(t) - phi(t)*theta(t)
So we need to move the line 69 before than parameter-updating (line 66). Is that right? or just I misunderstand it?

And I have another question. You may miss the updating of the F matrix. You may need to add the following code to line 44
self.F = (self.F - k @ hidden_vec @ self.F)/self.lbd ;

Thanks,

How to reproduce the results in table 1, 2 of the paper

Hello authors,

I'm trying to reproduce your results. I saw you produce some evaluation code in file data_analysis.ipynb but I cannot get the same results reported in the paper. Can you provide more detail about evaluation part?

Thank you.
I'm looking forward to hearing from you!