Kawato M. Developped the technique called the Minimum Commanded Torque
and Minimum Commanded Torque Change
to optimize joint torques (see article from Kaneko Y., Kawato M. and al., 2005).
A similar approach in the kinematics domain is the Minimum Jerk (MJ)
model. It have been proven to be the domain in which humans plan movements, see article from Flash T. and Hogan N., 1985 (not to mention it is also simpler to plan trajectories in the kinematics domain). Details of implementation can be found on Shadmehr R. website. It consist of 5th degree polynomials combined together and minimize the jerk (derivative of acceleration) over the whole trajectory, thus minimizing acceleration changes.
MJ in the articular coordinate system yield a unique trajectory in the articular space, that depends on the choice of the configuration for the robot at the intermediate passage points.
MJ model in the Cartesian coordinate system yield a unique trajectory in the Cartesian space, thus multiple possible solution in the articular space.
Using traditional optimization techniques with the MJ model is possible. Using the Resolved Motion Rate (RMR) technique after having computed the result of a pure MJ trajectory, the optimization function is
$z_{MJ}= \frac{1}{2}(q-q_{MJ})^TK^TK(q-q_{MJ})$
with $z_{MJ}$ the function to minimize, $K$ the gain coefficients, $q$ the articular position, and $q_{MJ}$ the MJ articular position. Computing the gradient $h$ of such function will allow you to find $\Delta \theta$ (see Huo, L. and Baron, L., 2008, for similar technique for joint limit avoidance).