-
Notifications
You must be signed in to change notification settings - Fork 1
Class Design
There are Eight classes defined in this system (so far)
Is a hollow class that is used to simulate an action agent applied to an environment. so far it has nothing useful except for its ID, which can be replaced by a simple integer I am thinking of adding extra features to this class
main player in this simulator, using a policy it learns by time, it interacts with the environment using possible actions, and recieves a signal with current state, previous state, reward, and action applied
container of the entire simulator. Each environemnt is made of:
- One World Dynamics object
- a set of agents
- a set of states
- a set of actions
Each agent has one policy defined. A policy will have:
- a state action pair value
- a state value
- a parameters object containing important parameters used by the policy policies can learn and act based on optimal values
an Event signal is a collection of objects used to declare the result of an action applied on the environment.
it contains:
- State after action
- Previous state before action
- Reward signal
- Action used to make this change
- Actor agent performing this action
A hollow class, similar to action. this is used to represent a state.
This is the main simulator
A new class to be defined, Action Collection, is a set of actions sorted in a certain manner, and can be accessed using an index. this index is wel understood by the agent related to this action collection. While an agent has a new attribute of AgentType. An Environment should define relation between agent types and action collections
- Make Agents interact only with integer actions
- Make only environments able to see action objects
- make action objects able to invoke real events
- World Dynamics should interact with the meaning of an action not with the ID of an action
- Actions and Agents should have a defined relation