I'm working on a problem using DDPG.
Is it possible to add some intelligence in the initialization phase, such that the convergence time is improved/shortened and local optima are avoided as much as possible?
For example, this may include assigning (higher) probabilities to (better) actions (in the action selection algorithm) at the start of an episode. This hopefully leads to the agent discovering and selecting "better" actions faster, rather than starting from more random ones. Or this won't work since the neural networks will just unlearn these initial values during the training process?
Also, with the above description, am I better off using Soft Actor-Critic?