This folder contains the experimental code and data for the paper: Relational program synthesis with numerical reasoning, . accepted at AAAI23.
SWI-Prolog (8.4.2 or above)
Clingo (5.5.0 or above)
pyswip (You must install pyswip from the master branch!)
pip install git+https://github.com/yuce/pyswip@master#egg=pyswipZ3 Python APIpip install z3-solver
Experimental data and experimental results are in the folder ilp-experiments/results/{task-name}.
To reproduce the experimental results (train and test), you can run python ilpexp.py {experiment-name}. You can change the number of cores (default 1) in this file, and the systems tested and their settings in ilp-experiments/ilpexp/experiment.py. Experiment names are in this file.
To modify experimental data, change the experimental data generator files in ilp-experiments/ilpexp/problem/{task-name}.
To use NumSynth with your own data, import your data in a new folder in ./numsyth/examples and run:
python ./numsyth/popper.py ./numsyth/examples/{your-folder-name} --numerical-reasoning
This experimental framework is based on https://github.com/logic-and-learning-lab/ilp-experiments.
NumSynth is an inductive logic programming (ILP) system. NumSynth learns programs with numerical values. NumSynth is based on Popper.
If you use NumSynth, please cite the paper:
Céline Hocquette and Andrew Cropper. Relational program synthesis with numerical reasoning.
As Popper, NumSynth requires three files as input:
More details about how to set up these files are provided in Popper's documentation.
By comparison with Popper, you must use the flag "--numerical-reasoning" to enable numerical reasoning. Moreover, the bias file for NumSynth may contain numerical predicates. These are predicates which arguments of which predicates may be bound to numerical symbols. NumSynth supports 4 built-in numerical predicates which are as follows:
numerical_pred(geq,2).
type(geq,(float, float)).
directions(geq,(in, out)).
numerical_pred(leq,2).
type(leq,(float, float)).
directions(leq,(in, out)).
numerical_pred(add,3).
type(add,(int, int, int)).
directions(add,(in,in,out)).
numerical_pred(mult,3).
type(mult,(int, int, int)).
directions(mult,(in,out,out)).
Supported types are 'int' or 'float'.
Please note that geq, leq, add and mult are reserved keywords, which may not be redefined in the background knowledge.
To constrain the search, you can specify bounds for the numerical variables as follows:
bounds(geq,1,(-100,100)).
bounds(leq,1,(-100,100)).
bounds(mult,1,(-100,100)).
bounds(add,1,(-100,100)).By default, the maximum number of numerical values in a clause is 2. You can change this setting by adding the following declaration in the bias file, where 'x' is the desired maximum number of numerical predicates per clause:
max_numeric(x).Alternatively, you can use the command line option --max-numeric as follows:
python ./numsynth/popper.py ./numsynth/examples/{your-folder-name} --max-numeric {x} --numerical-reasoning
The bias file declaration has priority over the command line option.
You also can specify a precision when learning with float numbers. The default precision is $$10^{-2}$$, you can change it by using the command line option --numerical-precision as follows:
python ./numsynth/popper.py ./numsynth/examples/{your-folder-name} --max-numeric {x} --numerical-reasoning 0.0001
Examples of tasks are provided in ./numeric/examples. Tasks which require numerical values have a subfolder name which starts with 'numeric'.
NymSynth also supports learning programs with magic values. It extends MagicPopper. For more information on learning programs with magic values, see Learning programs with magic values and the corresponding code.