Update botorch_and_ax.md (#71)

Summary:
Revised documentation / narrative on ax integration.
Pull Request resolved: #71

Reviewed By: Balandat

Differential Revision: D14994033

Pulled By: eytan

fbshipit-source-id: 45dd4a0ce44c3bd5870e4f74fec8affe90ae1b53

Author: eytan

docs/botorch_and_ax.md

@@ -3,41 +3,36 @@ id: botorch_and_ax
 title: Using botorch with Ax
 ---
 
-[Ax](https://github.com/facebook/Ax) is a platform for optimizing experiments.
+[Ax](https://github.com/facebook/Ax) is a platform for sequential
+experimentation.
 It relies on botorch for implementing Bayesian Optimization algorithms, but
-provides much higher-level APIs that make it easy and convenient to specify
-problems. It also comes with powerful metadata management, storage of results,
-different APIs that support a variety of use cases. Ax makes it convenient to
-use botorch in many standard Bayesian Optimization settings. Simply put, if
+provides higher-level APIs that make it easy and convenient to specify
+problems, visualize results, and benchmark new algorithms. It also comes with powerful metadata management, storage of results, and deployment-related APIs. Ax makes it convenient to
+use botorch in most standard Bayesian Optimization settings. Simply put, if
 botorch is the "un-framework", then Ax is the "framework".
 
-Ax provides a `BotorchModel` that is a sensible default for modeling and
-optimization, but that can easily be customized by specifying and passing in
-custom model constructors, acquisition functions, and optimization strategies.
-
+Ax provides a `BotorchModel` (**TODO**: cross-link to Ax documentation) that is a sensible default for modeling and
+optimization which can be customized by specifying and passing in
+bespoke model constructors, acquisition functions, and optimization strategies.
+This model bridge utilizes a number of built-in transformations (**TODO**: make sure these transformations are documented in Ax, and link to them here), such normalizing input spaces and outputs to ensure reasonable fitting of GPs.
 
 ## When to use botorch though Ax
 
-*Short answer:* If it's simple to use botorch through Ax for your problem, then
-use Ax. This should apply to most standard use cases.
+If it's simple to use botorch through Ax for your problem, then
+use Ax.  It dramatically reduces the amount of bookkeeping one needs to do as a Bayesian optimization researcher, such as keeping track of results, and transforming inputs and outputs to ranges that will ensure sensible handling in (G)PyTorch. The functionality provided by Ax should apply to most standard use cases.
 
-For instance, say you want to tinker around with some parameters of your botorch
-model (e.g. the kind of kernel), but leave the rest of the the Bayesian
+For instance, say you want to experiment with using a different kind of surrogate model, or a new type of acquisition function, but leave the rest of the the Bayesian
 Optimization loop untouched. It is then straightforward to plug your custom botorch
-model into Ax to take advantage of Ax's various loop control APIs,
+model or acquisition function into Ax to take advantage of Ax's various loop control APIs,
 as well as its powerful automated metadata management, data storage, etc. See the
 [Using a custom botorch model in Ax](../tutorials/custom_botorch_model_in_ax)
-tutorial for how to do this.
-
-**TODO:** Link to docs on Ax's transformations.
-
+tutorial for more on how ot do this.
 
 ## When not to use Ax
 
-If you're working in a non-standard setting, where you might deal with very
-high-dimensional feature or parameter spaces, or where the model fitting process
-requires interactive work, then using Ax may end up becoming cumbersome. In such
+If you're working in a non-standard setting, such as those with high-dimensional or structured feature or design spaces, or where the model fitting process requires interactive work, then using Ax may not be the best solution for you. In such
 a situation, you might be better off writing your own full Bayesian Optimization
-loop outside of Ax (hey, it's all just python after all...). The
-[q-Noisy Constrained EI](../tutorials/closed_loop_botorch_only) tutorial shows
-how this can be done.
+loop outside of Ax. The
+[q-Noisy Constrained EI](../tutorials/closed_loop_botorch_only) tutorial and
+[variational auto-encoder](../tutorials/vae_mnist) tutorial give examples of how
+this can be done.