Learned impact regularizer — offers a transfer learned alternative to → Impact regularizer (defined)

The two are presented back to back under the same section, explicitly as a fork: "Define an Impact Regularizer" is followed immediately by "An alternative, more flexible approach is to learn (rather than define) a generalized impact regularizer via training over many tasks" (concrete-problems, §"Learn an Impact Regularizer:", p. 5). The argument for switching is a claim about transfer: "side effects may be more similar across tasks than the main goal is," illustrated by a painting robot, a cleaning robot, and a factory control robot all wanting to avoid knocking over similar objects. Where the hand-defined impact regularizer requires a designer to correctly formalize "change to the environment" once per task, the learned version bets that this formalization is close to task-invariant and can be trained once and reused, an empirical wager the hand-defined version never has to make.