What is the difference between model and policy w.r.t reinforcement learning
Both definition seems to state they are mapping from states to actions then what is the difference or am i wrong ?
This articlce really sums it up for you:
What is Model-Based Reinforcement Learning?
To Model or Not to Model
“Model” is one of those terms that gets thrown around a lot in machine learning (and in scientific disciplines more generally), often with a relatively vague explanation of what we mean. Fortunately, in reinforcement learning, a model has a very specific meaning: it refers to the different dynamic states of an environment and how these states lead to a reward.
...The policy is whatever strategy you use to determine what action/direction to take based on your current state/location.
The overall outcome of Reinforcement learning (or any learning really) is to develop a policy, that is a series of behaviours or actions to take when presented with a specific domain.
The reinforcement factor is that you continually re-run the learning process based on the results of prior learning, effectively you apply the new policy and learn from the results to improve the policy.
In Model based Reinforcement learning we use a model to represent the environment or domain, this documents the facts, or states as well as the possible actions. By knowing certain facts the policies can target theses states and actions specifically in each repetition cycle, testing and improving the accuracy of the policy, just as it improves the quality of the model.
Another way to look at the two is that the model is a record or result of the prior learning, it is the updated view of the environment. The model deals in facts or assumed facts, based on past policy execution results, the model hold the records of past executions, this data can be used to approximate the outcomes of taking certain actions from specific states. The Policy is the actual learnings on the behaviours, where as the model is the facts that back up and confirm our learnings.
This diagram from the same article simplifies the relationship between model and policy in Reinforcement Learning:
- How can I filter and count objects in DJANGO template?
- Explanation for MVC pattern needed
- Where is the boundary for creating a table with types in the database?
- How to access URL helper from rails module
- How can I know if a table column is a foreign key in Laravel through Model?
- Saving a Newly Created Eloquent Model as a Job Property Prevents the Job from Being Handled
- How to call OneToOneField reverse relationship in template (django)?
- ValueError with Scikit-Learn
- Loading and using a codeigniter model from another model
- Is it possible to use rails image_tag from inside a model?
- django %(class)s does not get correctly replaced when used in m2m relationship to self in abstract model
- NetLogo Plots not being updated in the interface
- How can I change an object reference to a new object that has implemented the same protocol in swift?
- Django model inheritance: Create a subclass using existing super class
- AttributeError: The layer has never been called and thus has no defined output shape
- Hyper-prparameter tuning and classification algorithm comparation
- Joomla: two mysql queries model functions with one view
- Laravel ajax live search query with multiple models
- How to know the significance of coefficients and model in Weka?
- Is it bad practice to include non-validating methods in a pydantic model?
- Qt set a custom widget inside a QTableView
- Changing SwiftData @Model
- Model method not recognizing parameters
- Laravel: How do i fix this trait (with attributes) to also work when creating a model?
- Cannot overwrite model once compiled Mongoose
- How to solve the problems with empty return of model in view
- How to add a ForeignKey field to be saved in ModelForm Django?
- How to move model to the other section in Django's site admin
- Cloned trees are wrongly rotated
- Subset data into analysis and assessment data to generate and apply grouped prediction models in R