Respuesta :
Algorithms that use reinforcement learning can develop policies on their own for difficult jobs. However, the number of samples necessary to learn a wide range of abilities can be unreasonably high.
While meta-reinforcement learning techniques have made it possible for agents to use prior knowledge to adapt fast to new tasks, their success is critically dependent on how similar the new task is to the challenges they have already encountered.
The extrapolation capabilities of current techniques are either poor or can only be achieved at the cost of extremely high data requirements for on-policy meta-training. Model identification and experience relabeling (MIER), a meta-reinforcement learning technique that is effective and extrapolates well when confronted with out-of-distribution tasks at test time, is presented in this paper.
Our approach is founded on a straightforward realization: we understand that dynamics models, as opposed to policies and value functions, maybe more effectively and consistently changed using off-policy data. To continue training policies and value functions for out-of-distribution tasks without employing meta-reinforcement learning at all, these dynamics models can then be employed to create a synthetic experience for the new task.
Learn more about reinforcement learning: https://brainly.com/question/28462453
#SPJ4
