Use identical initial weights to better judge the convergence speed of different algorithms (i.e., with symmetrized loss or with prioritized experience replay)

[jakobatgithub]