Parallel Computing Toolbox deliberately sets the random number state of workers such that the each task of a job starts from a known state - i.e. Task 1 of every job starts from a particular state, and Task 2 of every job starts from another particular state etc. See this doc page for more.
So, in conclusion, the simplest approach is simply to place
rng('shuffle')
inside your script - but be warned that this approach is fragile (as mentioned in the doc page referenced above) because it is based on the time at which the code executes.
For a more robust approach, you need to pass in the random seed to your task, and set up the random state from that.
