minor fixes to RL and BNN codes

bayesian-code.ipynb

@@ -38,7 +38,8 @@
         "## Overview\n",
         "\n",
         "We are now considering the same setup as in the notebook {doc}`supervised-airfoils`: A turbulent airflow around wing profiles, for which we'd like to know the average motion\n",
-        "and pressure distribution around this airfoil for different Reynolds numbers and angles of attack. In the earlier notebook, we tackled this by completely bypassing any physical solver and instead training a neural network that learns the quantities of interest. Now, we want to extend this approach to the variational Bayesian Neural Networks (BNNs) of the previous section. In contrast to traditional networks, that learn a single point estimate for each weight value, BNNs aim at learning a _distribution_ over each weight parameter (e.g. a Gaussian with mean $\\mu$ and variance $\\sigma^{2}$). During a forward-pass, each parameter in the network is then sampled from its corresponding approximate posterior distribution $q_{\\phi}(\\theta)$. In that sense, the network parameters themselves are _random variables_ and each forward pass becomes _stochastic_, because for a given input the predictions will vary with every forward-pass. This allows to assess how _uncertain_ the network is: If the predictions vary a lot, we think that the network is uncertain about its output. "
+        "and pressure distribution around this airfoil for different Reynolds numbers and angles of attack. In the earlier notebook, we tackled this by completely bypassing any physical solver and instead training a neural network that learns the quantities of interest. Now, we want to extend this approach to the variational Bayesian Neural Networks (BNNs) of the previous section. In contrast to traditional networks, that learn a single point estimate for each weight value, BNNs aim at learning a _distribution_ over each weight parameter (e.g. a Gaussian with mean $\\mu$ and variance $\\sigma^{2}$). During a forward-pass, each parameter in the network is then sampled from its corresponding approximate posterior distribution $q_{\\phi}(\\theta)$. In that sense, the network parameters themselves are _random variables_ and each forward pass becomes _stochastic_, because for a given input the predictions will vary with every forward-pass. This allows to assess how _uncertain_ the network is: If the predictions vary a lot, we think that the network is uncertain about its output. ",
+		"[[run in colab]](https://colab.research.google.com/github/tum-pbs/pbdl-book/blob/main/bayesian-code.ipynb)"
       ]
     },
     {
@@ -1152,4 +1153,4 @@
       ]
     }
   ]
-}
+}