GaryGaulin
Posts: 5385 Joined: Oct. 2012
|
Quote (Cubist @ Aug. 06 2017,12:52) | Quote (GaryGaulin @ Aug. 06 2017,04:58) | Mudslinging creeps have made it worth staying out of this forum.
Management is setting a very bad example by condoning it. |
Gaulin. If you actually believed that this forum wasn't worth participating in… why in Babbage's name would you still participate in this forum? |
Good question.
At a time like this I dread having to say anything in this forum. Patrick Getty just moved to Texas for an apparently very nice job as professor. Even pays a living wage. That is sure a good sign that not all that is right with academia is lost. He will still be teaching what he most likes and says he we will be back next summer for more tracksite related work, but I worry about that not going as planned. It's like with all things considered in the area of education Texas has much more need for someone like him than Connecticut or Massachusetts has. I have to hope the best for wherever his next challenge leads. There remains something that is still going well with Holyoke Community College so it's not like nothing is locally happening, at the academic level.
Instead of dwelling on need for "participation" in such a lively exchange of ideas as are found filling this thread I have been pleasantly been writing code. For those who wondered where I went I just finished emailing this preliminary technical update worth sharing:
Quote | Camp, after over two weeks of trying different things I found a relatively easy way to get a ~98% average confidence level controlling a 4 speed Left/Right motor ramp where each increment doubles speed, without springing back and forth while following the spatial networks sometimes jittery direction vector. It has enough of a turn angle to build up momentum that takes almost half a turn to stop. But after a short time it still manages just fine.
I'm used to an ~75% confidence level. Past that usually indicates it's ready to get stuck at 100% by doing something that feels good but does not get it anywhere, sort of like a case of OCD. This time the high readings are because it's making all the right moves to become a very athletic dancer.
Memory required including in RAM addressing both before and after (two or three bit each) Speed and heading Error readings. Even tough the system can only think from one timecycle to the next this causes memory to temporally connect together (plan out) and smooth out the required time dependent actions.
It is possible to eliminate the time consuming challenges of the David Heiserman way of doing things by having the spatial network directly supply the Left/Right error correction signal. But without a body and enough mind to on its own figure out the muscle coordination details it's a poor representation of the real-world biology and physics. What made it hard to get it to reliably control logarithmic motion made it easy to develop spatial network behavior. There are now interconnection maps galore but how neurons are physically wired together has mostly lead to even more puzzling questions. Thankfully though the Heiserman way of modeling behavior made it easy for me to focus on possibilities such as cellular level brain wave generated vector maps where each place has the real-world properties of what is mapped there.
The only complication is that getting to 98% confidence takes 16 or so bits of memory space. For all it can do this is not all that much RAM. Unfortunately there is also a Forward/Reverse to same way control and my PC runs out of memory after reaching 28 bits. It's possible to take a few bits out of addressing without causing too much loss of performance, but I want 100% and can easily enough achieve that by making each a separate subsystem. Learning rate would also then be far greater, which more matters when controlling logarithmic motion.
In the past I experimented with connecting systems together with two way feedback, where it more or less figures out what to use the communication line for. I do not see there being a need for fancy crosstalk right now but it's another one of those things that makes modeling this way worth the occasional challenges.
Let me know whether what I'm coding has already been figured out and I just don't know where to look, or something. I'll otherwise keep working on what I described above then email you the resulting code so that you can see how bare-bones simple it is. There are no "2^LfRtErrWhatever" or other mathy calculations. At most multiplication and division, like to convert error in degrees to an index for an array that has the same log curve as the motor ramp array. This quickly finds the 0 to 3 number corresponding to that amount of error distance, and avoids equations that even puzzle me a year after coding them.
Gary |
This latest marathon project all started with my wife enjoying watching and critiquing the experimental ID Lab videos. After I had the automated recording code working real nice she then needed me to give the critter logarithmic left/right turning motion, as it somewhat had for forward/reverse. This forced me to give the software another good going over.
Now that Camp demonstrated the program for me where he lives out in California my starting work on what the robot builders most need right now was very timely. My finally finding an easy way to sort out a coding nightmare in the last ID Lab is another big worry now gone. Therefore instead of my wasting time arguing with Wesley I need to get back to work on biologically relevant dancing robots and related things.
-------------- The theory of intelligent design holds that certain features of the universe and of living things are best explained by an intelligent cause, not an undirected process such as natural selection.
|