<s>
In	O
machine	O
learning	O
,	O
a	O
hyperparameter	B-General_Concept
is	O
a	O
parameter	O
whose	O
value	O
is	O
used	O
to	O
control	O
the	O
learning	O
process	O
.	O
</s>
<s>
Hyperparameters	B-General_Concept
can	O
be	O
classified	O
as	O
model	O
hyperparameters	B-General_Concept
,	O
that	O
cannot	O
be	O
inferred	O
while	O
fitting	B-Algorithm
the	I-Algorithm
machine	I-Algorithm
to	I-Algorithm
the	I-Algorithm
training	I-Algorithm
set	I-Algorithm
because	O
they	O
refer	O
to	O
the	O
model	O
selection	O
task	O
,	O
or	O
algorithm	O
hyperparameters	B-General_Concept
,	O
that	O
in	O
principle	O
have	O
no	O
influence	O
on	O
the	O
performance	O
of	O
the	O
model	O
but	O
affect	O
the	O
speed	O
and	O
quality	O
of	O
the	O
learning	O
process	O
.	O
</s>
<s>
An	O
example	O
of	O
a	O
model	O
hyperparameter	B-General_Concept
is	O
the	O
topology	O
and	O
size	O
of	O
a	O
neural	O
network	O
.	O
</s>
<s>
Examples	O
of	O
algorithm	O
hyperparameters	B-General_Concept
are	O
learning	B-General_Concept
rate	I-General_Concept
and	O
batch	O
size	O
as	O
well	O
as	O
mini-batch	O
size	O
.	O
</s>
<s>
Different	O
model	O
training	O
algorithms	O
require	O
different	O
hyperparameters	B-General_Concept
,	O
some	O
simple	O
algorithms	O
(	O
such	O
as	O
ordinary	B-General_Concept
least	I-General_Concept
squares	I-General_Concept
regression	I-General_Concept
)	O
require	O
none	O
.	O
</s>
<s>
Given	O
these	O
hyperparameters	B-General_Concept
,	O
the	O
training	O
algorithm	O
learns	O
the	O
parameters	O
from	O
the	O
data	O
.	O
</s>
<s>
For	O
instance	O
,	O
LASSO	B-Language
is	O
an	O
algorithm	O
that	O
adds	O
a	O
regularization	O
hyperparameter	B-General_Concept
to	O
ordinary	B-General_Concept
least	I-General_Concept
squares	I-General_Concept
regression	I-General_Concept
,	O
which	O
has	O
to	O
be	O
set	O
before	O
estimating	O
the	O
parameters	O
through	O
the	O
training	O
algorithm	O
.	O
</s>
<s>
The	O
time	O
required	O
to	O
train	O
and	O
test	O
a	O
model	O
can	O
depend	O
upon	O
the	O
choice	O
of	O
its	O
hyperparameters	B-General_Concept
.	O
</s>
<s>
A	O
hyperparameter	B-General_Concept
is	O
usually	O
of	O
continuous	O
or	O
integer	O
type	O
,	O
leading	O
to	O
mixed-type	O
optimization	O
problems	O
.	O
</s>
<s>
The	O
existence	O
of	O
some	O
hyperparameters	B-General_Concept
is	O
conditional	O
upon	O
the	O
value	O
of	O
others	O
,	O
e.g.	O
</s>
<s>
Usually	O
,	O
but	O
not	O
always	O
,	O
hyperparameters	B-General_Concept
cannot	O
be	O
learned	O
using	O
well	O
known	O
gradient	O
based	O
methods	O
(	O
such	O
as	O
gradient	O
descent	O
,	O
LBFGS	O
)	O
-	O
which	O
are	O
commonly	O
employed	O
to	O
learn	O
parameters	O
.	O
</s>
<s>
These	O
hyperparameters	B-General_Concept
are	O
those	O
parameters	O
describing	O
a	O
model	O
representation	O
that	O
cannot	O
be	O
learned	O
by	O
common	O
optimization	O
methods	O
but	O
nonetheless	O
affect	O
the	O
loss	O
function	O
.	O
</s>
<s>
An	O
example	O
would	O
be	O
the	O
tolerance	O
hyperparameter	B-General_Concept
for	O
errors	O
in	O
support	O
vector	O
machines	O
.	O
</s>
<s>
Sometimes	O
,	O
hyperparameters	B-General_Concept
cannot	O
be	O
learned	O
from	O
the	O
training	O
data	O
because	O
they	O
aggressively	O
increase	O
the	O
capacity	O
of	O
a	O
model	O
and	O
can	O
push	O
the	O
loss	O
function	O
to	O
an	O
undesired	O
minimum	O
(	O
overfitting	O
to	O
,	O
and	O
picking	O
up	O
noise	O
in	O
the	O
data	O
)	O
,	O
as	O
opposed	O
to	O
correctly	O
mapping	O
the	O
richness	O
of	O
the	O
structure	O
in	O
the	O
data	O
.	O
</s>
<s>
Most	O
performance	O
variation	O
can	O
be	O
attributed	O
to	O
just	O
a	O
few	O
hyperparameters	B-General_Concept
.	O
</s>
<s>
The	O
tunability	O
of	O
an	O
algorithm	O
,	O
hyperparameter	B-General_Concept
,	O
or	O
interacting	O
hyperparameters	B-General_Concept
is	O
a	O
measure	O
of	O
how	O
much	O
performance	O
can	O
be	O
gained	O
by	O
tuning	O
it	O
.	O
</s>
<s>
For	O
an	O
LSTM	B-Algorithm
,	O
while	O
the	O
learning	B-General_Concept
rate	I-General_Concept
followed	O
by	O
the	O
network	O
size	O
are	O
its	O
most	O
crucial	O
hyperparameters	B-General_Concept
,	O
batching	O
and	O
momentum	O
have	O
no	O
significant	O
effect	O
on	O
its	O
performance	O
.	O
</s>
<s>
An	O
inherent	O
stochasticity	O
in	O
learning	O
directly	O
implies	O
that	O
the	O
empirical	O
hyperparameter	B-General_Concept
performance	O
is	O
not	O
necessarily	O
its	O
true	O
performance	O
.	O
</s>
<s>
Methods	O
that	O
are	O
not	O
robust	B-Application
to	O
simple	O
changes	O
in	O
hyperparameters	B-General_Concept
,	O
random	B-Algorithm
seeds	I-Algorithm
,	O
or	O
even	O
different	O
implementations	O
of	O
the	O
same	O
algorithm	O
cannot	O
be	O
integrated	O
into	O
mission	O
critical	O
control	O
systems	O
without	O
significant	O
simplification	O
and	O
robustification	O
.	O
</s>
<s>
Reinforcement	O
learning	O
algorithms	O
,	O
in	O
particular	O
,	O
require	O
measuring	O
their	O
performance	O
over	O
a	O
large	O
number	O
of	O
random	B-Algorithm
seeds	I-Algorithm
,	O
and	O
also	O
measuring	O
their	O
sensitivity	O
to	O
choices	O
of	O
hyperparameters	B-General_Concept
.	O
</s>
<s>
Their	O
evaluation	O
with	O
a	O
small	O
number	O
of	O
random	B-Algorithm
seeds	I-Algorithm
does	O
not	O
capture	O
performance	O
adequately	O
due	O
to	O
high	O
variance	O
.	O
</s>
<s>
DDPG	O
(	O
Deep	O
Deterministic	O
Policy	O
Gradient	O
)	O
,	O
are	O
more	O
sensitive	O
to	O
hyperparameter	B-General_Concept
choices	O
than	O
others	O
.	O
</s>
<s>
Hyperparameter	B-General_Concept
optimization	O
finds	O
a	O
tuple	O
of	O
hyperparameters	B-General_Concept
that	O
yields	O
an	O
optimal	O
model	O
which	O
minimizes	O
a	O
predefined	O
loss	O
function	O
on	O
given	O
test	O
data	O
.	O
</s>
<s>
The	O
objective	O
function	O
takes	O
a	O
tuple	O
of	O
hyperparameters	B-General_Concept
and	O
returns	O
the	O
associated	O
loss	O
.	O
</s>
<s>
Apart	O
from	O
tuning	B-General_Concept
hyperparameters	I-General_Concept
,	O
machine	O
learning	O
involves	O
storing	O
and	O
organizing	O
the	O
parameters	O
and	O
results	O
,	O
and	O
making	O
sure	O
they	O
are	O
reproducible	O
.	O
</s>
<s>
In	O
the	O
absence	O
of	O
a	O
robust	B-Application
infrastructure	O
for	O
this	O
purpose	O
,	O
research	O
code	O
often	O
evolves	O
quickly	O
and	O
compromises	O
essential	O
aspects	O
like	O
bookkeeping	O
and	O
reproducibility	O
.	O
</s>
<s>
Reproducibility	O
can	O
be	O
particularly	O
difficult	O
for	O
deep	B-Algorithm
learning	I-Algorithm
models	O
.	O
</s>
