Live Monitoring of Critic Predictions in the RL Toolbox

1 次查看（过去 30 天）
walli
walli 2020-8-17
编辑： walli 2020-8-17
I'm wondering if it is possible to monitor the Q-value predictions within any critic-based RL approach using the RL toolbox? For example, having a multi-output DQN agent the internal deep NN has to be called at every step in order to evaluate all possible discrete actions given the current state sample - hence, somewhere internally there must be a Q-value prediction for every discrete action available which are then evaluated in order to find the optimal action.
However, having spend some time on the 2020a documentation I was not able to find a way accessing these internal Q-value predictions at each time step. In particular, it would be nice if the Simulink-based agent block would be able to provide these predictions for further processing and monitoring reasons during the training and deployment phase.
Does somebody have a useful hint in order to retrieve the Q-value estimates during learning?

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 Reinforcement Learning 的更多信息

标签

产品

Community Treasure Hunt

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

Start Hunting!

Translated by