Paper-AC-Decisions-Learning

The project files for the paper 'Algebraic Circuits for Decision Theoretic Inference and Learning'. Submitted to ECAI2020. The main features are

• maxeu.py - Solves the maximum expected utility problem by constraining the ordering in the algebraic circuit (X-constrained SDDs).
• gradientmeu.py - Solves the maximum expected utility problem by utilising the algebraic circuit as function to be optimised. The circuit is a function which calculates the expected utility given a set of decisions parameters. These parameters are optimised via gradient ascent. To illustrate the power of algebraic circuits, the gradient is calculated using the circuit.
• ulearner.py - Learns a value for the unknown utility labels, using partially observed intepretations of which the total utility is observed.

More details on these approaches can be found in the paper.

Experiment files

script_dataset.py

A script used to create a ProbLog dataset from Bayesian networks. The Bayesian networks were first compiled into ProbLog programs using ProbLog's conversion script (12/2019). The script adds utilities and decisions, creates the input model and the partially observed examples to learning from.

script_maximisation.py

A script used to perform a maximisation experiment (constrained or unconstrained) on a created ProbLog BN dataset. The dataset creation relies on script_dataset.py and performs a maximisation experiment using maxeu.py or gradientmeu.py

script_maximisation_report.py

Small code example that was used to interactively extract information from previous maximisation experiments.

script_learning.py

A script used to perform a learning experiment on a created ProbLog BN dataset. The dataset creation relies on script_dataset.py and performs a learning experiment using ulearner.py

script_test.py

Small code example that was used to interactively extract information from previous learning experiments.

script_plot.py

A script used to create a plot (MSE, MSE_Test, RME vs epoch) for each previously executed learning experiment in a specified folder.

script_msse_plot.py

A script to calculate and create a plot of the mean squared sampled error (MSSE) over each epoch of the learning process.

script_scatter_plot.py

A script to create scatter plots (regret vs drop_prob) of regret.p files in a specified folder. The regret files can be created using evaluator.py

evaluator.py

A script to calculate the regret of a learned ProbLog model compared to the correct model.

Requirements

ProbLog 2.1.0.39+

PySDD 0.2.9+

Paper experiments

Maximisation

To reproduce the results of the paper, run the following:

script_maximisation.py survey --decisions 1 --approach both --seed 5 --repeat 10 --max_epoch 80 --lr 1.0
script_maximisation.py survey --decisions 2 --approach both --seed 5 --repeat 10 --max_epoch 80 --lr 1.0
script_maximisation.py asia --decisions 1 --approach both --seed 5 --repeat 10 --max_epoch 80 --lr 1.0
script_maximisation.py asia --decisions 2 --approach both --seed 5 --repeat 10 --max_epoch 80 --lr 1.0
script_maximisation.py earthquake --decisions 1 --approach both --seed 5 --repeat 10 --max_epoch 80 --lr 1.0
script_maximisation.py earthquake --decisions 2 --approach both --seed 5 --repeat 10 --max_epoch 80 --lr 1.0

Several lr's were tried, the conclusion remains roughly the same.

Report

script_maximisation_report.py

Learning

Perform experiments [asia, survey, earthquake] x [0.0, 0.1, ..., 0.9, 0.25, 0.75] x 5

script_learning.py survey --drop 0.0 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py survey --drop 0.1 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py survey --drop 0.2 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py survey --drop 0.25 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py survey --drop 0.3 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py survey --drop 0.4 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py survey --drop 0.5 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py survey --drop 0.6 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py survey --drop 0.7 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py survey --drop 0.75 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py survey --drop 0.8 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py survey --drop 0.9 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py asia --drop 0.0 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py asia --drop 0.1 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py asia --drop 0.2 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py asia --drop 0.25 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py asia --drop 0.3 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py asia --drop 0.4 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py asia --drop 0.5 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py asia --drop 0.6 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py asia --drop 0.7 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py asia --drop 0.75 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py asia --drop 0.8 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py asia --drop 0.9 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py earthquake --drop 0.0 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py earthquake --drop 0.1 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py earthquake --drop 0.2 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py earthquake --drop 0.25 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py earthquake --drop 0.3 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py earthquake --drop 0.4 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py earthquake --drop 0.5 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py earthquake --drop 0.6 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py earthquake --drop 0.7 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py earthquake --drop 0.75 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py earthquake --drop 0.8 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150
script_learning.py earthquake --drop 0.9 --seed 5 --max_epoch 80 --left_out "" --repeat 5 --nb_of_examples 150

Evaluating [survey1, ..., survey5] x [0.0, 0.1, ..., 0.9, 0.25, 0.75] x 3

evaluator.py survey --drop 0.0 --left_out "" --seed 5 --nb_of_examples 150 --nb_of_eval 3
evaluator.py survey --drop 0.1 --left_out "" --seed 5 --nb_of_examples 150 --nb_of_eval 3
evaluator.py survey --drop 0.2 --left_out "" --seed 5 --nb_of_examples 150 --nb_of_eval 3
evaluator.py survey --drop 0.25 --left_out "" --seed 5 --nb_of_examples 150 --nb_of_eval 3
evaluator.py survey --drop 0.3 --left_out "" --seed 5 --nb_of_examples 150 --nb_of_eval 3
evaluator.py survey --drop 0.4 --left_out "" --seed 5 --nb_of_examples 150 --nb_of_eval 3
evaluator.py survey --drop 0.5 --left_out "" --seed 5 --nb_of_examples 150 --nb_of_eval 3
evaluator.py survey --drop 0.6 --left_out "" --seed 5 --nb_of_examples 150 --nb_of_eval 3
evaluator.py survey --drop 0.7 --left_out "" --seed 5 --nb_of_examples 150 --nb_of_eval 3
evaluator.py survey --drop 0.75 --left_out "" --seed 5 --nb_of_examples 150 --nb_of_eval 3
evaluator.py survey --drop 0.8 --left_out "" --seed 5 --nb_of_examples 150 --nb_of_eval 3
evaluator.py survey --drop 0.9 --left_out "" --seed 5 --nb_of_examples 150 --nb_of_eval 3

Report

script_plot.py "0.7_0.5_150/"
script_plot.py "0.8_0.5_150/"
script_msse_plot.py survey --drop 0.7 -- seed 5 --nb_of_examples 150 --nb_of_samples 100
script_msse_plot.py survey --drop 0.8 -- seed 5 --nb_of_examples 150 --nb_of_samples 100
script_scatter_plot.py survey --left_out "" --nb_of_examples 150

Figure 4: ./data/results/0.8_0.5_150/survey_80_1337671202_False

Figure 5: ./data/results/0.7_0.5_150/survey_80_1337671202_False

Figure 6: ./data/results/scatter.txt

Figure 7: MSSE from each survey network used in learning