Python implementation of policy search and model training using $\alpha$-divergence minimization for Bayesian neural networks with latent variables. See:

Depeweg, Stefan, et al. "Learning and policy search in stochastic dynamical systems with bayesian neural networks." arXiv preprint arXiv:1605.07127 (2016).

Requires the standard libraries for theano-based models and Lasagne (I use 0.2.dev)

1. Insert industrialbenchmark_python in environment/:

   • Download python version of industrialbenchmark.

   • Move to environment/industrialbenchmark

2. Generate batch of state transitions:

    cd  environment/  
    python make_data.py 
   

   will generate a training and test set stored in environment/out

   X: Setpoint,A(t-14),..,A(t+1),R(t-15)..,R(t)

   Y: R(t+1)

3. Model Training:

   cd experiments/  
   python train_model.py 0.5 
   

   Will train a BNN using bb-alpha with alpha=0.5

   After training model will be stored in experiments/models as pickle file

   Code will run on GPU/CPU. Parameters are chosen conservatively for GPU use. Consider decreasing sample size to 25 for CPU use.

   Expected training time (i5-6600K CPU @ 4.0GHz, GTX 1060): CPU:
   50 samples: 21.5 hours
   25 samples: 10.5 hours

   GPU:
   50 samples: 3.5 hours
   25 samples: 2.0 hours

4. Policy Training

   cd experiments/  
   python train_controller.py 0.5 
   

   Will train a policy using model from step 2 (required a model exists in models/)
   After training the policy will be stored in experiments/controller as pickle file.

   Code will run on CPU. For GPU use one should pass only indexes to train_func using givens. In our experiments no speedup was obtaiend from GPU use.

5. Policy Evaluation:
   An example policy evaluation script is given in environment/eval_pol.py

   cd environment/  
   ipython  
   from eval_pol import evaluate   
   results = evalute('../experiments/controller/AD_1.0.p')  
   

Some helpful tips:

model.bb_alpha.network.update_randomness(n_samples)   

will sample n_samples from q(W) and resample the input noise.

For prediction use:

m,v = model.predict(np.tile(X,[n_samples,1,1]))  

where X is n x d
m is n_samples x n x d
v is n_samples x n x d (constant output noise variance)