Hey folks.
Over the past couple days, I wanted to focus on using PicoChess as my analysis tool. I've used it in a pretty extensive manner, with local engines, remote engines, remote tutor, connected to a DGT chessboard and with no e-board attached.
During all these tests, I've come across, a few times, some issues. The majority of the times the main issues I've encountered were:
- PicoChess becoming unresponsive, like it was hanging even if not 100% stuck;
- The "Analysis - Show" sometimes showing [E] when you would have expected to see [T] and vice versa;
- Erratic behaviour of the Tutor engine, compared to the "Turbo setup" recently introduced and/or Tutor active/inactive compared to the corresponding Tutor Active/Inactive toggle on the moder UI menu.
The point is that these erratic situations do not show up if you simply setup a position, scan it, launch analyse. In this case, with a very straightforward approach, everything seems to work fine.
The issues arise when you toggle the engines on and off a few times, when you setup a position, you scan it, then put it into Play Mode, then you enable Tutor, then you stop it... in other words... when you use it like you would use an analysis tool in real life during an analysis session :-)
In a nutshell, my own feeling is that "in principle" we have a machine that does work, but it's not yet robust nor reliable enough. Don't get me wrong. We have an absolutely great tool in our hands, and I couldn't be more enthisiastic and with the fondest appreciation for those who have spent all their time and energies in it.
I'm only saying the we still need to fix a few things. And thankfully we have, so we can continue enjoy with our creature :-D
That said, I'm now into trying to document a test sequence that would generate some of those issues, a replicable path to provide you a trace to follow and fix it.
I'll try and come back to this thread when I find one.
Thanks.
AG