Skip to main content

Models

Explanation​

A model is the ADK's approach to the Don't Repeat Yourself principle. Models solve the problem of having to develop / copy-paste the same agent across different directories for each system being optimized. Instead, the agent can be implemented once and the system-specific details can be part of the model.

This separates the concerns between developing a generic algorithm / method to optimize systems (for an agent), and tuning specific details to optimize a specific system (for a model).

Creating a model​

In the ADK's eyes, a model is a directory within an agent's models directory with the following contents:

  • models/: A directory containing agent specific model files (NN checkpoints, stats, etc.), completely controlled by the developer.

  • hyper_parameters.json: A file containing the hyper-parameters for the agent's training / the hyper-parameters the agent was trained with. See Hyper Parameters for more info.

  • metadata.json: A file containing the model metadata. See Metadata for more info.

  • target_specifications.json: A file containing the target specifications that can be defined for the model. See Model TargetSpec for more info.

  • world_control_specifications.json: A file containing the world controls (controllable parameters of the system) that the model can optimize for. See World Control Specifications for more info.

Run genie model add in an agent directory and follow the instructions in order to conveniently create a model for that agent.

note

The adk package has to be installed before it and its related tools such as genie can be used. It is highly recommended that one uses a python venv for installing the ADK.

What next?​

Once an agent with the above structure has been successfully created either manually or through genie model add, start the agent. While the ADK is running, the model sync service registers, updates, and removes models on the platform when local files under the models directory change.

If a model is registered on the platform and the agent is running, a compatiable model will be capable of being selected for an optimization on the platform.

The model sub-command for genie provides a convenient interface for adding, listing, deleting, and restoring agent models. Run genie model -h for more info.

For transfer, export, and import between models, see Model handling. For how models fit into optimization runs, see What To Do Next.