general value/q network for multi-dimension reward

[PaParaZz1]

Motivation

There are some cases about multi-dimension rewards, both for environments and algorithms:

• some complicated env, such as auto-drive, speed, stability, collision and other elements can lead to some kinds of reward
• some algorithm designs, such as intrinsic reward in exploration related algorithm

But If we want to use multi value/q network to learn different rewards, we need to do some non-trivial modifications in current DI-engine policy, so we need a general design and validate its necessary in performance.

Plan

TODO