Hello Everyone,
I am recently working on the topic " Deep Reinforcement Learning in production scheduling".
Most recently I am working on the simulation of the environment, the state and action engineering process.
The uprising question for me here is: how are uncertainty and constraints taken into account in a RL project?
In stochastic optimization there is a probability distribution of several factors which are taken into account. Also the solution space is restricted with constraints.
How are these factors considered in model-free RL? Is the uncertainty included in the simulation of the environment? Are the state or action space restricted through constraints?
Best regards,
Christoph
Dear Christoph,
when you use a model-free reinforcement learning algorithm like Q-learning, you do not need to model the environment but also determine the action-space and state-space. I recommend you to use MushroomRl library or gym library in python
to simulate your environment. for more information, you could review these documents:
MushroomRL — MushroomRL 1.7.0 documentation
Making a custom environment in gym | by Ashish Poddar | Medium
Eventually, the model will learn uncertainty after interacting with the environment and adapt to that.
Please note that we might not know the probabilities in the environment before we start. We can develop them by averaging while playing.
The learning rate parameter can be adjusted according to our interest.
The question is do we want to explore or exploit the environment??
- Badges
- Science topic