- 0.3.1 - Testing completely different inputs and improvede movement. Much improved movement.
- 0.3 - Trying a really shallow input history and Mako style movement.
- 0.2 - Tweaked the NN and its inputs. For better or worse, I have no idea. Now saves and restores (huge) NN files.
- 0.1 - The first version. Numbered with a 0 since it is just something I threw together in a jiffy and I will need to think it through much more before it can be called 1.0.
What's special about it?
It's my first robot with NeuralTargeting
. I'm using NRLIBJ
Where can I download it from?
How does it move?
How does it fire?
, crude neural targeting, but still. I will need to experiment a lot with what inputs to give the network and what topology to use in the hidden layers. Right now I just feed it some enemy history and such and the output is a fireing angle relative to the current absolute bearing of the enemy. This doesn't work very well. I'm not sure if it can be made to work, but my intuition whispers that I am doing it in a too simple way. But like Tad
I'm not giving up easily on things. =) Besides pitching Orca up against Marshmallow
shows that it must be working somewhat, Orca gets 13.8% hitrate and wins in both survival and Robocode score. I think M might be the only top20 bot that Orca can beat though. Quite stressing a fact...
How does the melee strategy differ from one-on-one strategy?
This is strictly a OneOnOne
What does it save between rounds and matches?
It saves the neural networks between round. And between matches as well.
Where did you get the name?
I figured Orca is an intelligent creature. My Orca is probably an insult to real Orcas...
Can I use your code?
Not at the moment. I will eventually release the code for all my bots, but not while I am still competing. I'm more than happy to share ideas though.
What's next for your robot?
- Make it work.
- Figure out how I can make it learn a bit faster.
What other robot(s) is it based on?
Please ask questions about and/or comment Mako
Be careful with network biases. I found unable to predict bearings, but I think it can be because the network doesn't uses biases by default. You can change it from the NrPop? class. -- Albert
It seems Orca's ability to predict bearings is limited to simple movent like SpinBot, Walls, Calliope and such. Ares type of mixed mode movements totally tricks Orca's NN. What is this about biases? -- PEZ
Anybody want's to help me tune this targeting? I'm running out of ideas on how to make it work. -- PEZ
A "neuron" recieves an input and then passes it to a non-linear function. The general formula is:
input = w0 + w1*x1 + w2*x2 + ... + wn*xn
If you don't set the biases, w0 will always be 0. Some times you need to have them, some other not. -- Albert