Policies

Policies determine the agent’s behavior by mapping observations to actions. Policies can be deterministic (e.g. greedy policies) or stochastic (e.g. exploratory policies). A policy can also be learned by an agent, or it can be just a simple rule-based policy (e.g. a fixed building schedule).

In this module, we provide two classes. SchedulePolicy is used to define rule-based policies that can be used as baselines for RL agents. SavedModelPolicy is a wrapper used to load and interact with policies that have been learned and saved.

All of the policies in this project should tf_agents.policies.TFPolicy to provide a consistent interface to interact with policies.

SchedulePolicy

Purpose: Implements traditional rule-based control strategies based on time schedules.

Implementation: SchedulePolicy in schedule_policy.py

Usage Example: see the implementation of the create_baseline_schedule_policy function in policies/schedule_policy.py, which instantiates an example baseline policy.

SavedModelPolicy

Purpose: Loads and uses a previously trained policy that has been saved.

Implementation: SavedModelPolicy in saved_model_policy.py

Usage Example:

from smart_control.reinforcement_learning.policies.saved_model_policy import SavedModelPolicy

# Load a saved policy
policy = SavedModelPolicy(
    saved_model_path="path/to/saved/policy",
    time_step_spec=time_step_spec, # obtained through tf_env.time_step_spec()
    action_spec=action_spec # obtained through tf_env.action_spec()
)

---

Back to Reinforcement Learning

Back to Home