Subject:
|
Re: Very strange problems
|
Newsgroups:
|
lugnet.robotics.rcx.legos
|
Date:
|
Sun, 21 May 2000 18:28:46 GMT
|
Viewed:
|
1240 times
|
| |
|
|
> > We haven't reached the GA part yet. Right now the idea is to work out a
> > communication protocol and then send a information back and fourth.
>
> Please make sure to share this with us! IR continues to be legOS's
> weakest point, and once someone works out a reliable way to do this,
> it'll be extremely useful for many people.
We have not done anything to ensure that whatever is transmitted is
received correctly (i.e. without errors). What we have done is a build a
system which will work even if one (or several) of the RCXs are
temporarily blocked for some reason.
When a genome is transmitted this is done ten bytes at a time and after
that a question is sent if all ten bytes were received. We send each
byte seperatly, but we later found out that one could send entire
arrays. We haven't tried this yet, though.
> > And
> > since the RCXs locked up we thought that there must be something wrong in
> > the programming.
>
> Unfortunately, no (though if you are multi-threading, there may be some
> issues there...)
Nope, no multi-threading.
> > It seems we just have to deal with the fact that legOS is strange and try
> > to get it to work anyway. I just wondered if we could be resonable sure
> > that a program that worked didn't just stop working. That wouldn't be good
> > after several hours of training the GA population :).
>
> Yes. We seriously considered doing GA stuff, but this was a concern. I'd
> suggest working out a backup protocol, by which the values of your genes
> (what is the technical term for that, anyway? I forget.) are regularly
I use this word :).
> transferred to the PC. That would give you something to fall back on if
> there is a problem.
> > > BTW, the neural net code is very slow, since it requires floating point
> > > emulation.
> >
> > We figure our GA system will also be pretty slow due to the fact that the
> > individuals need to be tested by actually moving the robot. Depending on
> > how fast we can transmit data over IR, this might be a time factor also.
>
> Oof. That will be painfully slow, I'm afraid. What problem are you trying
> to solve, exactly? I'm very curious, now :)
We plan to build a six legged robot with one RCX connected to each leg. We
will have a shared population with probably a max length of 100. The idea
is that the robot should learn to walk using the 30-monkies-in-a-bus
algorithm:
Four individuals are selected randomly.
Each individual gets to "drive" for some specified time. We will let each
individual run between two and four times and set the leg in the position
we get as output. The leg can be set either up or down and, using a
rotation sensor, be rotated to a specific direction.
The individuals recieve fitness. We will have a tailing wheel conncected
to a rotation sensor to register movement.
The two worst monkies are thrown off the bus (deleted) and the other two
are used to create two new monkies using cross-over and mutations.
Four new monkies are put behind the wheel and the algorithm runs again.
We figure that running one individual four times and setting the motors
every time will take about 30 seconds. Transmitting the genome will take
about the same as it looks now.
The neighboring legs position is used as input for the algorithm and each
individual is run on all legs simultaneously.
Tobias
|
|
Message has 1 Reply:
 | | Re: Very strange problems
|
| (...) Redundancy is good :) (...) Good. It's generally solid, but can cause some issues from time to time. <snip lots of monkey talk :) > (...) Distance of movement or consistency (i.e. always forward) or both for fitness? Sounds really cool. Good (...) (24 years ago, 21-May-00, to lugnet.robotics.rcx.legos)
|
Message is in Reply To:
| | Re: Very strange problems
|
| (...) Argh. I was getting emulegos (which I was pretty familiar with) and legosim (which I though was C, not Java) confused. I'm afraid neither of them would be too easy to port to 0.2.x :( (...) Please make sure to share this with us! IR continues (...) (24 years ago, 21-May-00, to lugnet.robotics.rcx.legos)
|
20 Messages in This Thread:
  
- Entire Thread on One Page:
- Nested:
All | Brief | Compact | Dots
Linear:
All | Brief | Compact
This Message and its Replies on One Page:
- Nested:
All | Brief | Compact | Dots
Linear:
All | Brief | Compact
|
|
|
|
