Decide whether MPC baselines should use hard VPP constraints or remain purely objective-driven

[xudongwu-0]

Problem

The current MPC baselines are primarily objective-driven. This keeps the method clean, but it may allow practically undesirable actions, such as scheduling controllable appliances during VPP events if the objective tradeoff allows it.

We need to decide whether the MPC baseline should remain a pure finite-horizon objective optimizer or include engineering hard constraints.

Open Design Question

Should mpc_dynamic and mpc_ep enforce hard constraints such as:

• no controllable appliance operation inside VPP windows
• water-heater preheat must end before VPP start
• EV charging should avoid VPP unless required for departure SOC
• shiftable tasks should either finish before VPP or start after VPP
• setpoint changes should respect persona-specific comfort boundaries while still reacting to VPP events

Tradeoff

Pure objective-driven MPC:

• Cleaner academically
• Easier to explain as finite-horizon cumulative-cost MPC
• Useful for studying how much the objective formulation alone can achieve
• But may produce bad edge cases if weights are imperfect

Hard-constrained MPC:

• More robust and practical
• Better VPP compliance
• Easier to compare against Agent behavior in deployment-like settings
• But less pure and adds engineering assumptions that need to be documented

Possible Resolution

Create an explicit baseline design decision:

1. Keep mpc_dynamic and mpc_ep as objective-only baselines.
2. Introduce separate constrained variants, such as mpc_dynamic_constrained and mpc_ep_constrained.
3. Report both objective-only and constrained MPC results when comparing against the Agent.

Related Questions

• Should hard constraints be applied only during active VPP windows, or also during pre-event planning windows?
• Should constraints come from the appliance simulator, the MPC candidate generator, or the objective scorer?
• How should unavoidable service constraints be represented, such as EV departure SOC or water-heater bath readiness?
• Should constrained MPC use the same role-play user scoring pipeline as the Agent baseline?