Noise model in RL for large action signal

1 次查看（过去 30 天）
SomeMatlabUser
SomeMatlabUser 2021-8-26
编辑： SomeMatlabUser 2021-8-26
I want to train a model with an DDPG agent. The model requires an action 10 element vetor signal with a bound value of -1.5...+1.5. The problem is, that the noise model seems to only change 5 elements per episode, while the other signals stay at the upper or lower limits. The noise is added to all signals over time, but it seems that only approx. 5 signals are changed at once. Example action signal that I get:
[1.5 0.2 0.4 -0.1 -1.5 -1.5 1.5 0.2 0.4 -0.1 -1.5]
If the noise does never change all signal elements together, the system will never find the spot, at which the system becomes stable. I let the training run for over 24h, without any change of the noise behavior. Is this a known thing and is there a way to alter the noise behavior? It did try to change the NoiseOptions parameters, but this does not change the way, how the noise affects the general problem I have, which still keeps always approx. 5 signals at their respective upper or lower bounds.

请先登录，再回答此问题。

回答（0 个）

类别

在 Help Center 和 File Exchange 中查找有关 Introduction to Installation and Licensing 的更多信息

标签

产品


版本

R2019a

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by