janweinreich / rules_of_acquisition

Bayesian Optimization made simple. Name derived from Star Trek universe, the "Rules of Acquisition" are a collection of sacred business proverbs of the ultra-capitalist race known as the Ferengi.

• 👋 Hi, I’m @janweinreich
• 👀 I’m interested in Machine Learning predictions of molecular properties & privacy preserving computation
• 🌱 I’m currently improving my skills for transformer neural networks
• 📫 [email protected]

Scholar 🎓

https://scholar.google.com/citations?user=3fhtth0AAAAJ&hl=en

Twitter 🐦

https://twitter.com/WeinreichJan

Homepage 🏡

https://janwein.com/

rules_of_acquisition's Issues

Enhancements

• devise an automatic stopping criterion for the BO based on expected improvement (ask alex)
• add the xgBoost regressor (with hyperparamteroptimization)
• qEI (without noise averaging)
• OHE versus random search comparison
• full entropy search (no gibbon) (Multi-Objective Max-value Entropy Search Acquisition Function)
• Curate old examples with 1 and 0 costs

elif scale_type_X == "botorch": not require scaling?

fix the default scaling options to be consistent

Code cleanup

Need to finally clean the code. Main aspects:

• remove unncesserary keywords from exp_configs.py for the AcqFctPrice script (for instance maximal price argument is not needed). This also means adding a surrogate argument to the surrogate="GP", in the BO.py

• specify regressor and acquisition function combinations

• clean the gibbon_search_modified_all function as it can now use different acquisition functions and so its name is not appropriate anymore

Budget planning

Given the current code for buying ligands and updating the prices accordingly,
implement a simple way that allows saving money first before buying new ligands, instead of a fixed budget (or even saving money from the last iteration).

The goal is to answer the question:
"Should I really first spend little or directly go for ligands with high promise according to the acquisition function?"

This question is hard to answer and it probably depends on many factors. Lets simply try to answer this question for the given dataset and not make general claims

New plans for experiments

basically I think that for the initialization and the budget based approach we need some more specific possible configs

"dataset": ["dolye2018", "dolye2022", "dolye2022_prices", "jensen2019"] # something like that
"init_strategy": ["strat1", "strat2", ...] # this is the most tricky one, probably need subconfigs, or at least a separate class for each, 'ntrain' should be in there too
"nruns": int
"niter": int # I'm thinking maybe we could add some other termination criterions but for now it should be fine
"batch_size": int
"strategy": ["random","normal_BO", "budget", "acq_score_per_price"] # there should also be subconfigs for 'budget' including 'max_batch_cost', 'fixed' price or 'increasing', w savings or wo, etc...  

missing experiment implementations

• FixBatch buchwald!

    {
        "dataset": "buchwald",
        "init_strategy": "worst_ligand_and_more",
        "cost_aware": True,
        "n_runs": 5,
        "n_iter": 60,
        "batch_size": 5,
        "max_batch_cost": 20.0,
        "ntrain": 200,
        "prices": "update_all_when_bought",
        "surrogate": "GP",
        "acq_func": "GIBBON",
        "buget_schedule": "adaptive_2",
    },    

current costmin.py there only works for direct acryklation

Include budget in acquisition function

Include the costs of ligands in the acquisition function by dividing the acqftvalue by the cost of the ligand. In case the ligand was already bought dont devide by zero but find a sensible lower bound