hi, zixuan, I trained the model using your codes on "Item Placement", and evaluated the rewards of 10, 20, 30, 40, 50 iterations and choose models that have top-3 rewards. But the reward on the test dataset is only 4400, I wonder if there are any other tricks I haven't added. Also, I found that the seed is not set in the script, does this affect the final result?

python generate_data.py item_placement --file_count ${i} --njobs 5 --train_size 10000 --valid_size 4000

Hello, I am master's student of A.I. course.

First, i'm really congratulations your ML4CO dual task winning.

I read your paper "ML4CO-KIDA: Knowledge Inheritance in Dataset Aggregation" recently. but i am beginner of this field(solving COP with ML), i have some questions for this paper.

1. Are there reasons why KIDA works better than EWA?(What i thought is, KIDA can calculate average policy evenly)

2. How can KIDA work well? I thought it will have some limitations, because it use strong branching as data which did bad performance on test case, and it trains with loss(strong branching, policy prediction)

3. in Figure 3: step 0~1000, there is big drop. is that coincidence? or representation of drop?(represent train accuracy at step 0)

hello, i found that the efficiency of training data generation is a big problem when running your codes. Although multi-threading is used here, as far as I know, python multi-threading is not much faster than single threading. So, i want to know if you have any other methods to speed up the generation of training data? Thank you！