resilient backpropagation parameters selection In the original paper for resilient backpropagation (http://paginas.fe.up.pt/~ee02162/dissertacao/RPROP%20paper.pdf), the author says "One of the main advantages of RPROP lies in the fact,
that for many problems no choice of parameters is needed
at all to obtain optimal or at least nearly optimal convergence
time".
Resilient backprop is described as a better alternative to standard backprop and adaptive learning backprop (in which we have to set learning rate and momentum).
No mention of setting the learning rate and momentum in resilient backprop is found in the paper mentioned above. I just wanted to confirm that do we need to set the learning rate and momentum for resilient backprop? 
 A: RPROP is certainly a nice algorithm for backpropogation.  It does avoid some parameters but it still requires some depending upon the library you are using (sometimes they are hidden with default values).  Essentially you are looking at the following variables all of which are noted in section C, page 588 of the paper you provided.  


*

*learningrate - yes, this still must be set at an initial value.  A good value is 0.1 as noted for a starting point.  A noteworthy point is that the $\eta = 0.1$ is actually a vector of multiple 0.1 equal to the length of your number of weights.  The reason people may say that you don't need to care about learningrate with RPROP is because it is going to be modified throughout the training process (similar to adaptive methods).  That said, you may find faster convergence in some situation with different initial learningrate values.

*learningrate limits - to prevent underflow/overflow of update values you set max and minimum values.  Good default values are $\Delta^{max} = 50$ and $\Delta^{min} = 1e-6$ respectively.

*learningrate factors - these determine how much of a jump the update values will take in either the positive or negative direction.  The recommended values are $\eta^- = 0.5$ and $\eta^+ = 1.2$.
Although RPROP is a nice algorithm, it still doesn't solve the hyperparameter problem entirely but with some solid default values you should be good for most circumstances.
EDIT
The learning rate component of the RPROP algorithm has been noted as confusing so here is my attempt to clarify.  This is ultimately a consequence of differing notation and terminology.  The learningrate is actually updated in equation (4) on page 587 of the paper linked in OP's question and reproduced below.
$
  \Delta_{ij}^{(t)} =
\begin{cases}
\eta^{+} * \Delta_{ij}^{(t-1)},  & \text{if $\frac{\delta E^{(t - 1)}}{\delta w_{ij}} * \frac{\delta E^{(t)}}{\delta w_{ij}} > 0$} \\
\eta^{-} * \Delta_{ij}^{(t-1)}, & \text{if $\frac{\delta E^{(t - 1)}}{\delta w_{ij}} * \frac{\delta E^{(t)}}{\delta w_{ij}} < 0$} \\
\Delta_{ij}^{(t-1)}, & else
\end{cases}
$
where $0 < \eta^{-} < 1 < \eta^{+}$
In this case, the learningrate(s) is actually not denoted as $\eta$ but actually the $\Delta_{ij}$ values.  This is because the learningrate is not a single value for the RPROP algorithm as noted above in my point 1.  It contains an update-value for each weight (as per the syntax in the paper).  These values are updated (hence an adaptive learning method) with each iteration.
Although the notation differs, the learningrate is still used in the formulas.  I hope this clarifies this point for other future readers.
