What do the ANOVA bars mean?

The ANOVA bars indicates which variables are important for the computation of the response. The statistical evidence is weak for the portion of the bar shown in red. If the whole bar is red, then you cannot trust the value computed, and you may be better of believing that that variable does not contribute to the response. Conversely, if most of the bar is gray, then the contribution of that variable is significant. Note that the ANOVA value is the gray and red part together; the red part is therefore half of the confidence interval and the other half of the confidence interval extending to the right of the ANOVA value is not shown.

The underlying technology considers b and b*b to be two different variables in the case of response surfaces. So, if you have y = 1.0 * b + 0.0 * b*b, then you will only see a bar for b and the b*b bar should be of length zero. Conversely for y = 0.0*b + 1.0*b*b, you will see a bar for b*b and the bar for "b" will be of length zero. So if b*b does not contribute much, you can change the order of your response surface from quadratic to linear. Note that you only have b together with b*b for response surfaces; for neural networks, you will only have b.