Check evaluation, possibly a bug
Closed this issue · 2 comments
Motor babbling is almost always better evaluate than goal babbling, whereas the latter covers the sensory space much better
Actually it seems to be normal.
In the motor babbling case, half of the testing points are generally reached perfectly (those in front of the arm origin), whereas the other half has a mean error around 0.2, hence a total mean error around 0.1. In the goal babbling case, this is rather 0 and 0.1, hence a total mean error around 0.05.
In summary, goal babbling is only slighty better than motor babbling at the end, but generally has a much lower standard deviation of errors.
A solution to have more significant result would be a arm environment where the points reached in motor babbling correspond to much less than a half of the total reachable space. I didn't find the appropriate configuration yet ...
In general it is better to use mean SQUARED errors. This penalizes large errors. In the example you mention, the different would be larger.
On 15 May 2014, at 13:41, Clément Moulin-Frier notifications@github.com wrote:
Actually it seems to be normal.
In the motor babbling case, half of the testing points are generally reached perfectly (those in front of the arm origin), whereas the other half has a mean error around 0.2, hence a total mean error around 0.1. In the goal babbling case, this is rather 0 and 0.1, hence a total mean error around 0.05.
In summary, goal babbling is only slighty better than motor babbling at the end, but generally has a much lower standard deviation of errors.
A solution to have more significant result would be a arm environment where the points reached in motor babbling correspond to much less than a half of the total reachable space. I didn't find the appropriate configuration yet ...—
Reply to this email directly or view it on GitHub.