Error should be squared

[toontalk]

The notebook states:

The error is

𝑅+𝛾max 𝑄̂(𝑠′,𝑎′;𝑤−)−𝑄(𝑠,𝑎;𝑤)

Which I couldn’t make sense of. Then later in the notebook the algorithm image has the above (with Q-hat) squared which makes sense since the sign of the difference shouldn’t matter. And the programming instructions refer to mean squared error.

[toontalk]

On further thought I find the use of the word “error” confusing. Gradient descent reduces the “error” but here it is reducing the square of the difference between the target and the predictions. So isn’t it clearer to refer to the error as (target-prediction)**2 instead of just target-prediction?