Cotton Baseline

This section compares three different fertilization policies used in an experiment to improve crop yield.

• The first policy is the “null” policy, which involves no fertilizer application, and is used as a reference point to measure the effect of nitrogen fertilizer on crop yield.

• The second policy is the “expert” policy, which is the fertilizer application policy used in the original experiment.

• The third policy is a “PPO” policy, learned with default hyper-parameters.

Note

The PPO policy was trained for \(10^6\) episodes with stochastic weather generation, and the best model was saved based on performance in a validation environment. The performance of each policy was measured with \(10^3\) episodes in a test environment; featuring different stochastic weather series from those used in training and evaluation environments of PPO, to avoid over-optimistic performance measures.

Training and evaluation code can be found in the following repository

Mean cumulated return of each of the 3 policies against the day of the simulation. Shaded area displays the [0.05, 0.95] quantile range for each policy.

2D histogram of fertilizer applications (the darker the more frequent).

Grain Yield Experiment Results

Policy	Grain Yield (kg/ha)	Percent nitrogen in leaf tissue (%)	Total Fertilization (kg/ha)	Application Number	Nitrogen Use Efficiency (kg/kg)
Null	365.1 (121.9)	1 (0.1)	0 (0)	0 (0)	n.a.
Expert	3199.4 (269.1)	1.4 (0)	133 (0)	4 (0)	21.3 (1.3)
PPO	2743.5 (341.9)	1.3 (0.1)	92.1 (16.4)	7.6 (2.1)	26.1 (2.3)