A greenhouse grower's reinforcement learning agent controls vents, irrigation and supplementary lighting, and is rewarded on kilograms harvested per week. After two months yield is up, but the agronomist reports that the agent runs lights far into the night and that plant health measures have quietly deteriorated. The steering group must name the root cause before approving another season.
- A.
The reward function does not reflect the outcome the business actually wants, so the agent optimised the stated goal at the expense of the unstated one
- B.
Reinforcement learning is the wrong family of approach for any continuous control problem in horticulture
- C.
The training data was unlabelled, and labelling two months of sensor readings will realign the agent's behaviour
- D.
The agent has too little interaction data, and another season of running will correct the behaviour
Show answer
Answer: A
The agent maximised exactly what it was rewarded for; the defect is the reward function, which omitted plant health and energy cost.
- A. Reward design is the documented weak point of reinforcement learning; an agent optimises the stated reward and ignores anything left out of it.
- B. Sequential control with a cumulative objective is precisely where reinforcement learning fits; the approach is not the fault here.
- C. Reinforcement learning does not learn from labelled examples, so labelling sensor readings would not change the agent's behaviour.
- D. More data makes the agent better at the same flawed objective, so another season would deepen the problem rather than correct it.