NERO mod enemy sensors actually see friends during training

[GoogleCodeExporter]

NERO mod enemy sensors are initialized with the wrong bit mask, making them see 
friends instead. This leads to a lot of sensor confusion and difficulty in 
early learning. This is a significant problem.

Original issue reported on code.google.com by ikarpov on 19 Nov 2011 at 6:52

[GoogleCodeExporter]

Original comment by ikarpov on 19 Nov 2011 at 6:52

• Changed state: Started

[GoogleCodeExporter]

Attached files, tnero-sensors and onero-sensors, illustrate the problem.

these are log-scale histograms of the network inputs accumulated over the 
course of learning to go to the flag.

as can be seen in onero-sensors, enemy sensors are activated despite there 
being no enemies on the field.

Original comment by ikarpov on 19 Nov 2011 at 6:57

Attachments:

• tnero-sensors.jpg
• onero-sensors.jpg

[GoogleCodeExporter]

[deleted comment]

[GoogleCodeExporter]

r1422 should address this issue

Original comment by ikarpov on 19 Nov 2011 at 7:08

• Changed state: Fixed

[GoogleCodeExporter]

this is the histogram set that verifies that the new sensors have been fixed.

Original comment by ikarpov on 19 Nov 2011 at 7:53

• Changed state: Verified

Attachments:

• onero-fixed-sensors.jpg

[GoogleCodeExporter]

further fixes were needed

Original comment by ikarpov on 21 Nov 2011 at 6:47

• Changed state: Started

[GoogleCodeExporter]

an uninitialized value was causing some sensors to return non-zero when zero 
values were required

additionally, the input layer of the network was not large enough to accept the 
bias, and was not causing an error

after r1431, this issue should be fixed

attached files, onero-fixed2 and onero-fixed3, show the sensor histograms with 
flag only and with flag and walls respectively.

Original comment by ikarpov on 21 Nov 2011 at 6:49

• Changed state: Fixed

Attachments:

• onero-fixed2-sensors.png
• onero-fixed3-sensors.png