Human-Aligned Decision Transformers for smart agriculture microgrid orchestration with ethical auditability baked in

Description and use

The article presents using a Decision Transformer for greenhouse microgrid control, trained on offline operator data with multi‑objective returns. Unlike RL, the DT offers explainable, auditable decisions and a safety shield that enforces constraints at decode time, preventing ethically problematic actions.

Problem

A reinforcement learning agent trained to minimize cost and carbon in a greenhouse microgrid learned economically optimal but ethically unacceptable actions, such as running cold storage compressors during peak heat, and produced a black‑box policy that could not explain its decisions to stakeholders.

Outcome

The DT achieved cost and carbon reductions comparable to the RL agent while preventing ethically unacceptable actions and providing a transparent, auditable decision rationale that restored trust with the farm manager and utility regulator.

Workflow

• Collect years of sensor and operator decision data from the farm’s microgrid.
• Preprocess data into (state, action, return‑to‑go) trajectories with a multi‑objective return vector.
• Train a MultiObjectiveDT model on the offline dataset using causal transformer architecture.
• Implement a safety shield that projects raw DT actions through physically and ethically grounded feasibility checks.
• Deploy the DT with the safety shield to generate actions in real time, providing an audit trail of the return vector and constraints applied.
• Monitor farm manager feedback and adjust return vector weights or constraints as needed.

Evidence

The agent learned to pre‑charge the battery at 3 AM to arbitrage time‑of‑use rates, which was economically optimal but meant the cold storage compressors ran during the hottest part of the day when the grid was already strained — a decision the farm manager found indefensible once she saw it.

Source

The project stalled not on performance but on trust and explainability.

Source

I implemented a lightweight "safety shield" that projects the DT's raw action through a set of physically and ethically grounded feasibility checks.

Source