<div class="gmail_quote">On Wed, Aug 17, 2011 at 07:14, Richard Katz <span dir="ltr"><<a href="mailto:richard.katz@earth.ox.ac.uk">richard.katz@earth.ox.ac.uk</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
<div id=":yo">For my uses this is a very important feature.<br>
<br>
Two questions:<br>
<br>
- Will the step size increase after a shorter step has been successful? Will it go immediately back to the maximum size?<br></div></blockquote><div><br></div><div>The details depend on the adaptive controller. Unfortunately, we haven't yet unified the interface for adaptive controllers, so, for example, TSALPHA and TSGL have different APIs. But both of them provide adaptive controllers now and the rate of increase can be limited.</div>
<div><br></div><div>I think having these interfaces is really undesirable and that we should unify it, but the information provided by the error estimates for TSGL are quite different from most methods (OTOH, they seem to me noisy which limits their utility, but there are other possible reasons for that, including "starting methods").</div>
<div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;"><div id=":yo">
<br>
- Will it be possible to set a minimum step size, below which the TSSolve() or TSStep() will "crap out?"<br>
<br>
It is obviously important that TS not just chug along without error if a timestep fails.</div></blockquote></div><br><div>So this is possible, but it's somewhat inconsistent between methods. It would help a great deal to build a suite of test problems that provide some way to evaluate error (even just through self-convergence) so that we can test whether an adaptive method is performing well.</div>
<div><br></div><div>My intent is to provide some common interface for adaptive controllers as well as some sample controllers. The controller can evaluate whether to accept or reject a step as well as choosing the next time step and, for the variable-order families, the method to use for the next step (selected from a list of candidate schemes). This is basically what I did in TSGL and later, and what Lisandro did in TSAlpha, but we need to unify the interface despite these methods giving us somewhat different information in their (embedded or extrapolated) error estimates.</div>
<div><br></div><div>I think TSAlpha has the best tested adaptive controller right now. Getting all the cool adaptive features into TSARKIMEX is next, hopefully with a unified interface for user-provided controllers.</div>