ICNN2SP

Implementation of the method proposed in the paper "ICNN-enhanced 2SP: Leveraging input convex neural networks for solving two-stage stochastic programming" (https://arxiv.org/abs/2505.05261)

About

ICNN-enhanced 2SP leverages input convex neural network (ICNN) to exploit the linear programming (LP) representability in solving two-stage stochastic programming (2SP) problems. The codebase extends Neur2SP’s implementation (https://github.com/khalil-research/Neur2SP) by augmenting it with ICNN-specific functionality.

Setup and Usage

The pre-generated datasets for benchmark problems are included in the data folder. The best neural network parameters found through random search for each problem instance and surrogate model type are stored in data/nn_params.py.

ICNN-enhanced approach

To use the ICNN feature, run the commands below sequentially:

# Step 1: Train the ICNN surrogate model
python runner.py --problems {PROBLEMS} --train_icnn_e 1

# Step 2: Obtain the best model
python runner.py --problems {PROBLEMS} --get_best_icnn_e_model 1

# Step 3: Reformulate 2SP and obtain solutions
python runner.py --problems {PROBLEMS} --eval_icnn_e 1

Replace {PROBLEMS} with any of the following problem instances:

• cflp_10_10, cflp_25_25, cflp_50_50
• sslp_5_25, sslp_10_50, sslp_15_45
• ip_b_E, ip_b_H, ip_i_E, ip_i_H

MIP-based approach (from Neur2SP)

To use the MIP-based method to solve 2SP (the original Neur2SP approach), simply replace icnn with nn in the command parameters (i.e., use --train_nn_e, --get_best_nn_e_model, and --eval_nn_e) and follow the same steps as above. This trains a conventional ReLU neural network surrogate using the same dataset as the ICNN-enhanced method, but reformulates the 2SP problem as a mixed-integer program.

Baseline evaluation

To obtain the baseline results by evaluating extensive form, run:

python runner.py --problems {PROBLEMS} --eval_ef 1 --n_cpus {N_CPUS}