Welcome to Dr.
Watson, part detective (think Sherlock Holmes), and part
psychologist (think John B. Watson , one of the founders of
modern psychology). Dr. Watson is an interactive tool designed to
help scientists guide laboratory experiments. At this time, it is
oriented primarily toward choosing continuous parameters
(concentration, temperature, pressure, time,...) required to run
We assume experiments are run
sequentially (we see the results of one experiment before running
the next), and we assume that the experiments take enough time
that it is worth thinking about what to do next. We further
assume that you are trying to maximize or minimize some metric
The tool at this
stage is extremely young, with limited functionality. It has been
tested primarily using Safari on a Mac, and Chrome on Windows. It
does not work at this time with Internet Explorer.
Dr. Watson starts with a Bayesian prior, which means that we work
with what you already know. As of this writing, we have to assume
that the relationship between your performance metric (say,
strength) is a quadratic function of each input parameter (such
as temperature or concentration). But we assume you do not know
which quadratic function. As the tool evolves, we will be adding
Dr. Watson walks you
through a series of steps:
Step 1: Belief
expression - You use a graphical toolkit to draw a family of
Step 2: Prior
formation - Using the family of curves that you have drawn, Dr.
Watson creates a mathematical representation of a prior using
normal distributions. We create a best estimate, and can show you
a set of samples drawn from a distribution we create from the
information provided in Step 1.
Step 3: Value
of information - We use a tool called the knowledge gradient to
identify the next experiment that is likely to produce the
highest possible improvement. We show the range of outcomes (in
the form of updated beliefs known as the posterior) that would
result if you ran one more experiment using parameter settings
that you specify.
Step 4: Posterior belief -
This is the updated prior after running an experiment. You can
run an experiment with a high value of information, or another
one that you choose.
Much more to come, but
that is all we have developed so far.