Mouse-GYM Training Protocols

A Mouse-GYM Training Protocol is a mechanism for incremental training through one or more phases of device configuration and system behavior. Progress within a protocol is tracked individually for each animal. Animals may have multiple protocol histories.

Training Protocols

A Training Protocol is the definition of the specific procedures and requirements to move through phases of training. It is defined once. The progress of individual animals with a protocol is tracked separately. A protocol consists of:

• id a unique identifier for each protocol generated by the system (not guaranteed to be user-friendly)
• name a user-friendly name/identifier
• description more detailed information than the name (optional)
• phases a set of Training Phases (defined later) to progress through

Training Phases

A Training Phase defines the state of the system for given phase of training (which can be modified during the phase) and requirements for exiting the phase. One set of requirements represents the condition(s) to advance to the next phase if there is one. A second, optional set of requirements represent the condition(s) to fall back to the previous phase due to insufficient progress in some amount of time or any other supported metric.

At the end of every behavior trial, three steps are taken for the current phase:

• Any defined "trial actions" are called. These do not directly affect exiting the phase.
  • An example could be moving the pellet arm after every trial where a pellet is consumed.
  • A trial action could modify a property of the training progress that would affect the phase exit behavior to follow
  • These actions, if present, are also called at the start of the phase
• If fall-back conditions are defined, they are evaluated and if the result is True, the system transitions to the previous phase
• If advance conditions are defined, they are evaluated and if the result is True, the system transitions to the next phase

Trial actions, fall-back conditions, and advance conditions are all optional.

The definition of what is a separate training phase vs. something that is modified during trial actions (e.g., changing pellet delivery location) is up to the Training Protocol writer.

A Training Phase definition consists of:

• id a unique identifier for each phase generated by the system (not guaranteed to be user-friendly)
• name a user-friendly name/identifier
• description more detailed information than the name (optional)
• fallback condition condition(s) for returning to the previous phase
• advance condition condition(s) for advancing to the next phase
• trial actions predefined actions or arbitrary Python code to execute after each trial

There are also default settings for most behavior algorithm settings, such as:

• pellet delivery enabled
• pellet covering enabled
• baseline magnet intensity
• reach distance to uncover
• auto-clamp enabled
• auto-clamp release delay
• auto-clamp load count

These are set when the phase is entered. They may be changed by trial actions.

Training Predicates

Training Predicates define the conditions for exiting a phase, either to advance or fall back. There are built-in predicates for simple tests and a Python predicate that will evaluate user-defined Python code. Predicates include:

• Number Comparison will test <, <=, ==, >=, or > for integer or floating point values
• String Comparison will test string values
• Python will execute user-defined python code (not fully implemented)
• Compound can be set to return True if all child predicates are True or set to return True if any child predicates are True (OR vs AND)

Compound predicates can be used to assemble any combination of any or all conditions at any depth.

Predicates have access to all public behavior algorithm, tunnel device, and pellet device public properties, as well as all properties tracked as progress for the animal such as pellets consumed. Properties at any depth in these objects can be reached via property paths delimited by periods such as progress.pellets_consumed.

Training Actions

Training Actions are performed each time a trial completes. They can be used to make adjustments during a phase that do not constitute moving a new phase.

Similar to predicates, there are predefined actions that do not require user-defined Python code to execute. If none of the predefined actions provide the required functionality, user-defined Python code can be used. These include:

• HeadMagnetIntensityAction - defines a range of values (start, increment, end) of head magnet intensity values to progress through based on the specified number of pellets successfully consumed for each step
• ReachDistanceBasicAction - defines a range of values (start, increment, end) of pellet-y offset values to progress through based on the specified number of pellets successfully consumed for each step
• Python - will execute user-defined python code (not fully implemented)

Training actions may also contribute to determining whether a phase is complete. In addition to checking predicates, if an action is marked as has_progress, its is_complete property will be evaluated after each trial.

Training Progress

Training Progress is the per-animal tracking of progress through the protocol and each phase. This is primarily used to evaluate exit conditions. Values include:

• Amount of time since the phase was first entered (calendar time, i.e., Sept. 12, even if the system was not running the entire time)
• Active time in phase - the amount of time actually spend with the phase active and system running
• Pellets presented
• Pellets consumed
• Successful reaches
• Starting pellet delivery location for this phase
• Current pellet delivery location (might be different if modified by actions)

This a partial list and the list will grow as more metrics are required for assessing progress or reporting.

There is also a generic dictionary available for Python trial actions and predicates to read and write any JSON serializable value.

Notes for Code Development

• The user-facing Training Protocol entity is represented in the class TrainingPlan due to specific use of Protocol in Python.