Francisco Coelho / zugzwang

notes.md 11.2 KB

Zugzwang Meetings

2024-04-17 - Online meeting

  • Anchor in the "0000" refs:
    • Annotated Logic Programming
    • Need to investigate how/if our semantic for annotations is compatible.
    • What about update from "observations" and the "score" of a program?
    • Probabilistic Database
    • Need to investigate how to incorporate the combination strategies for interdependence.
  • Example suggestion?

2024-04-16 - Presencial meeting

  • Abandon KR, consider EPIA.
  • Keep the ICLP chance.
  • Anchor in refs "0000".
  • Give more examples.
  • Explain better what we've done and what we've used..

2024-03-25 - Meeting

  • Próxima reunião: terça, 11:00 + \epsilon
  • Reestruturação.
  • Todos vêm "Thorough comparison with related work"
  • Tratar "Give a general argument [about Bayesian networks]"
  • Secção 4 para apêndice?
  • ICLP -> foco na linguagem

2024-03-25 - Work on KR24

From "Twelve Definitions..."

  • ... programs in which recursion and negation "don't mix." Such programs are called stratified.
  • stable model semantics, as well as the well-founded semantics can be seen as an attempt to generalize and simplify the iterated fixpoint semantics of stratified semantics.
  • three theories of nonmonotonic reasoning
    • circunspection: syntactic transformation that turns a first-order sentence F into the conjunction of F with another formula, which expresses a minimality condition.
    • default logic: I didn't got it
    • autoepistemic logic: modal operator Lp: "p is believed".

2024-03-20 - Work on KR24 and ICLP24

  • s(cASP): installed, tested, performance dismaying (can't solve 8-queens) vs. clingo that solves 8-queens in 0.004s and 100-queens in
  • Detail syntax and semantics

2024-03-15 - IJCAR24 Reviews

Summary

  • State-of-the-art:
    • Thorough comparison with related work
  • Motivation:
    • Clarify the application of the approach
    • Explore the advantages and limitations of the formalism
  • Technical details:
    • Self-containment
    • Detail syntax and semantics of the considered class of programs.
    • Clarify the relation of stable models and events
    • Recall the stable model semantics and its properties
    • Argument for Proposition 1 [is not] convincing
  • Fixes:
    • Provide the probabilities of the classes and of the events
    • Clarify the role of "testing of the prior distributions"
    • Give a general argument [about Bayesian networks] instead of an illustration on a simple example.

See Reviews file.

2024-01-30 - Exploratory Research Project

Apply for FCT funding.

2024-01-05 - Next Research Lines

After the base-setting work of "An Algebraic Approach to Stochastic ASP" these are the next tasks to consider. Is summary:

  1. Logic Programming - Stratified & Non-stratified programs
  2. Computer Science - Inductive Logic Programming
  3. Software - Integration with Potassco and other frameworks
  4. Applications

Line 1: Logic Programming - Stratified & Non-stratified programs

Line 1a

Stratified & non-stratified programs are quoted in the "CREDAL" papers as important classes of logic programs.

Minimal example of a non-stratified program.

The following annotated LP, with clauses $c_1, c_2, c_3$ respectively, is non-stratified (because has a cycle with negated arcs) but no head is disjunctive:

0.3::a.                 % c1
b :- not c, not a.      % c2
c :- not b.             % c3

This program has three stable models: $$ \begin{aligned} m_1 &= \set{ a, c } \cr m_2 &= \set{ \neg a, b } \cr m_3 &= \set{ \neg a, c } \end{aligned} $$

We should investigate What are stratified programs and why are they important? and how does our approach deals with such programs?

Line 1b - Investigate the expressiveness of PASP

Consider:

  • Recursion
  • Variables,
  • functional symbols,
Line 1c - The equivalence relation

Consider the cases where only $s \subseteq e$ and $e \subseteq s$. Or other refinements. Also consider the inconsistent and independent events.

Line 1d - Stability of the error function

Consider alternative error functions. See statistics, Kullback-Leibler divergence

Line 2: Computer Science - Inductive Logic Programming

Proceed from scoring programs to support genetic algorithms or other program space exploration methods.

Scoring programs, as described in our paper, is just a step into Inductive Logic Programming. To go further, we need to explore algorithms that:

  1. Use background knowledge, expressed as a PLP.
  2. Consult positive examples that should be soft induced.
  3. Consult negative examples that should be soft excluded.
  4. Generate PLPs that are scored.
  5. Recombine the best scored into a new population, using recombination rules.

In order to do that, PLPs must be expressed as data structures to be manipulated. Also recombination rules must investigated before become formally expressed and supported with adequate methods.

Line 3: Software - Integration with Potassco and other frameworks

Support annotated programs with zugzwang semantics.

  • Bayesian Networks (BII Alice)
    • Generate an annotated asp program from a bayesian network and run it trough clingo.
    • Recover the stable models from the previous ste and compute the respective probabilities.
  • Program Manipulation
    • Annotated ASP program representation and a parser.

Line 4: Applications

Apply zugzwang to a few showcases, besides the theoretic corner stones (non-stratified, disjunctive, bayes networks), preferably based in real world scenarios, with complex structure and large datasets.

  • (Stochastic) Plan Generation
  • Yale-Shooting Problem
  • (Stochastic) Situation Calculus
  • Frame Problem
  • Latent Facts - and core assumptions.
  • Given a Bayesian Network (or a Markov Networks):
    • Represent it. (done for BNs; MNs?)
    • Solve the common probability tasks: join (done), marginals, conditionals, parameter learning, inferring unobserved variables, sample generation, etc.
  • Given a solved ASP specification:
    • What is the marginal probability of the atom a? (done)
    • What other probability queries are important to consider?
  • Given an unsolved ASP specification:
    • What is the probability (distribution?) of the probabilistic fact a?
    • What other questions are relevant? E.g. the distribution family of a fact?
  • Given a solved ASP specification and a set of samples:
    • How do the probabilities inferred from the specification match the ones from the empiric distribution? (done might see alternative approaches)
  • Given two solved ASP specification and a set of samples:
    • Which specification best describes the empiric distribution? (done)

2024-01-05 - Publish Paper "AASASP"

Target conferences to publish paper "AASASP" Shared spreadsheet: https://docs.google.com/spreadsheets/d/1KJdA-zj3Xly-_tSE6tjbd7M9SttC9xBixXoqWZJFVfg

Conference Abstract Deadline Conference Date Location OBS
IJCAR 2024 2024-01-29 2024-07-3:6 Nancy, France ~Picked~ Rejected
ECAI'24 2024-04-19 2024-10-19:24 Santiago de Compostela, Spain
KR 2024 2024-04-24 2024-11-2:8 Hanoi, Vietnam
GECCO 24 2024-02-05 2024-07-14:18 Melboune, Australia Overdue
ICLP 24 2024-04-15 preferred
JELIA 25
ICFP 24 2024-03-01 2024-09-2:7 Milan, Italy

2023-02-28 - Looking for Application Examples

What applications are we looking for?

  • (Stochastic) Plan Generation
  • Yale-Shooting Problem
  • (Stochastic) Situation Calculus
  • Frame Problem
  • Given a Bayesian Network (or a Markov Networks):
    • Represent it.
    • Solve the common probability tasks: marginals, conditionals, parameter learning, inferring unobserved variables, sample generation, etc.
  • Given a solved ASP specification:
    • What is the marginal probability of the atom a?
    • What other probability queries are important to consider?
  • Given an unsolved ASP specification:
    • What is the probability (distribution?) of the probabilistic fact a?
    • What other questions are relevant? E.g. the distribution family of a fact?
  • Given a solved ASP specification and a set of samples:
    • How do the probabilities inferred from the specification match the ones from the empiric distribution?
  • Given two solved ASP specification and a set of samples:
    • Which specification best describes the empiric distribution?

What should be the task for the scholarship student? Use the Python API of clingo.

  1. Read a string and extract probability annotations; Associate those annotations with the respective atoms.
  2. Call clingo to get stable models.
  3. Support computation of the equivalence classes: Which functions and relations?
  4. Compute event probability using weighted model counting on the equivalence classes.
  5. Read a Bayesian Network from a file (BIF, DSC, NET, RDA, RDS, ...) and generate an annotated "ASP" specification.

2022 - AAAI - Inference and Learning with Model Uncertainty in Probabilistic Logic Programs

  • Is "Epistemic Uncertainty (EU)" the right framework for Zugzwang? How relevant are the epistemic questions in this paper to our work?
  • EU can be represented by Credal Sets, Subjective Logic and Beta Distributions?
  • Experiments made with BNs from (Kaplan and Ivanovska 2018) and larger networks from the BNLearn repository.
  • Are networks, Bayesian Networks in particular, a "good enough" pool of "example applications" to us, for now?

2023-01-10 - 15:00

  • Paper
  • Project
  • Latent Facts

2022-12-12

  • Is the project proposal ok? How long/detailed should it be?
  • Initial exploratory code event_lattice.py and EventLattice.ipynb done.
  • Start writing paper: Introduction, state of the art, motivation
    • Identify key problems
    • Target Conferences
    • KR;
    • ICLP;
    • ECAI
  • Next task for prototype:
    • Get stable models from potassco/s(casp)
    • other?

2022-12-05

Resources

Conferences