#gecco2023 — Public Fediverse posts

Fun facts from the closing notes:
- on-site attendants are ~57 x more heavy in CO2 emissions over 5 days than online, making all attempts to limit emissions outside travel all but meaningless

49/🧵 #GECCO2023 #GECCO

23. Black/Gray/White Box applications:
- High-intensity applications:
- Spiking NNs to data
- Mapping protein energy landscapes

24. Mix all of the above & Evolve by interacting with other communities.

48/🧵 #GECCO2023 #GECCO

19. Hybrid systems:
- Neuroevolution
- Evo-optimized gradient descent

20. Use of cultural artifacts instead of population:
- Hall of Fame of populations
- Archives of populations
- Multi-run memory
- auto-restarts allowing to learn from previous runs (meta-learning)

21. Landscapes analysis
- Landscape categories
- Exploit landscape structures
- Meta-EA for landscapes exploitations
- Adaptive EAs for Landscapes

22. "No Free Lunch"
- Algo/Problem fit

47/🧵 #GECCO2023 #GECCO

17. Rant about communication: Generative Learning is sexy, but not the EA flavor:

=> Tbh, the sexy part of generative learning is the same as was for CovNets; it's the fact that it works, and it works outside academia; which is not the case for EA-based solutions

18. Time-varying environment for the optimization problems: adaptation and tracking.

15.2. Evolving Agent-based models:
- Robotics
- Cyber-Sec
- Viral Evolution
=> that's Mut. Landscapes, not EC

46/🧵 #GECCO2023 #GECCO

13. Practitioner-friendliness:
- Design
- Parallelization:
(=> Consensus FTW)
- Isolated islands to be taken advantage on
- finely grained/cellular models
- Asynchrony
(=> Consensus FTW)

14. Co-evolutionary settings:
- Arms race not achieved yet (GANs failed)
- Instead a transation towards solving Min-Max Problems
- Cooperative co-evolution

15. Agent-based models with evolution

16. Transition from parameters to design spaces to design patterns

45/🧵 #GECCO2023 #GECCO

8. Unification: Evolutionary Computation book by Kenneth De Jong; definition of EAs and domains as algorithms instantiated from heuristics

9. Transition from named to property-base EAs

10. EAs got too complicated. Algorithms got explosively complicate due to meta-parameters, no visibility for eg biologists;

11. Evolution towards robust EAs; to self-adapting EAs;

12. Automating EA design itself => meta/hyper-heuristics

44/🧵 #GECCO2023 #GECCO

3: 1970s: Parallel evolution of the three independent branches.

4: 1980s: Continued parallel development; first conferences

5. 1990s: Cross-breeding starts between domains. Intergroup interactions; common field name emerges: **Evolationary Computation**; first journals: EC + IEEE Trans. on EC

6. emergence of GECCO/CEC/PPNS; first Dagstuhl meetings

7. 2000s: Institutional ties: ACM/IEEE + Cambrian explosion, both wrt problems and algorithms diversity

43/🧵 #GECCO2023 #GECCO

Time for the last session of the conference - SIGEVO keynote by Kenneth De Jong: "Evolutionary Computation Evolving".

1. Roots of Evolutionary computing: 1950s. First ANNs, EAs; First annealing, summarized in "Evo computation; The Fossil Record"

=> Darwinian vision of evolution

Vision that complexity can emerge from simple rules (Santa-Fe institute).

2. 1960s: Speciation:
- Evolutionary Strategies
- Evolutionary Programming
- Genetic Algorithms

42/🧵 #GECCO2023 #GECCO

Time for a "Hot Off the Press" session!

6. Improving symbolic regression in multi-objective scenarios by switching fitness functions and learning morphology of expressions

7. Improving Hypervolume optimization in multi-objective optimization by using hypervolume approximation - invariant network to find non-dominated solution sets.

41/🧵 #GECCO2023 #GECCO

Time for a "Hot Off the Press" session!

1. Maintaining diversity of solutions in multi-objective problems

2. Using concepts from theory of evolution to improve Siamese networks performance in multi-class optimization

3. Find optimal recombination strategy when Genetic Programming chromosome feature combinations are sparse

4. Bi-objective optimization tracking by using wavelets and superpositions

5. Replacing population sampling to Hermite-Gaussian quadrature

40/🧵 #GECCO2023 #GECCO

6. Optimization of fast and certifiably correct assembly code for cryptography, matching human code and beating the compilers.

7. LLMs X Genetic programming to generate autonomous code by doing a feedback on the bugs detected in the code.

=> IMHO testing programs is the hard task here, especially for non-trivial toy tasks for which tests can be auto-generated.

8. Granulates-based logic gate design to beat Moore's law with artificial evolution.

39/🧵 #GECCO2023 #GECCO

HUMIES Sessions:

1. Optimizing algorithm to prioritize ambulance calls.

2. MAGE: Detection of patterns in malware code with EAs

3. Fluid mechanics for particle-laden flows: improving symbolic rules for simulation relative
to human-written rules for simulations.

4. Predictive forecasting of hypoglicemic events based on measures: glucose level, HR, teps, Calories, ...

5. Interpretable models to identify rare and under-diagnosed diseases with computable incl. criteria

38/🧵 #GECCO2023 #GECCO

"To slide or not to slide? Moving along fitness levels and preserving the gene subsets diversity in modern evolutionary computation"

Looks like a plateau navigation theory - more specifically the NK model of evolution, which is useful for zero-initialized problems that are common in the optimization.

Examples on Max3SAT of 2 algorithms augmented with sliding - LT-GOMEA; P3; 3LOa.

37/🧵 #GECCO2023 #GECCO

"The impact of asynchrony on parallel model-based EAs"

The question is whether we can drop the synchrony to gain performance.

=> To me, it looks like crashing resilience, which is weaker than byzantine-resilience.

Evaluation time biases is the thing that is known to work on that.

Focuses on GOMEA algorithm with a linkage tree; with a GOM. It is sequential

Tbh, it really looks like a re-discovery of crash-resistant consensus from distributed computing...

36/🧵 #GECCO2023 #GECCO

Uses JAX to run it in massively parallel fashion.

Synthetic data on loss landscapes shows good reslts.

For neuroevolution tasks, even
in small populations (<10) in few iterations (<100) gives decent results.

Parent x Children rank => multi-parent/children pairs replaced at the same time; allowing elitism to be tuned in.

Found new algorithms can be transferred to other GAs.

LES: Learned Evolution Strategies.

Overall looks really interesting.

36/🧵 #GECCO2023 #GECCO

"Discovering Attention-Based Genetic Algorithms via Meta-Black-Box Optimization."

Software 2.0 - Move from manually-written algorithms to learning algorithms..

The idea is to evolve algorithms that are better in search thanks to hyper-parameter optimization.

Featurize => Cross-Attention (with layer skipping) => Select; in order to get rules as to whether a child would replace the parents or now; then perform mutation adaptation to optimize the sampling.

35/🧵 #GECCO2023 #GECCO

This transfers to learning dependencies in general additive functions.

Unfortunately, the best performing algorithm is rather slow.

The acceleration seems to work by doing memoization and transforming the problem into a tree representation.

34/🧵 #GECCO2023 #GECCO

Time for another session - Genetic Algorithms; starting with "First Improvement Hill Climber with Linkage Learning - on Introducing Dark Gray-Box Optimization into Statistical Linkage Learning Genetic Algorithms".

Looks like the tree of linkage of genes that interact as a tree, eg to do cross-over masks.

Apparently, false positives on linkage detection is a problem.

Recent algorithms to detect linkage seem to be free of such false-positives.

33/🧵 #GECCO2023 #GECCO

The change is to switch the program to AST and encode it with metamorphic encoding and record only AST edge mutations, eg reverse conditions; add an unused variable, ....

The next step is to derive a fitness metric from it and perform an Evolutionary Search on it.

Evaluation on 350 programs sent to Code2Vec seem to show that the search is effective and efficient.

32/🧵 #GECCO2023 #GECCO

"Searching for Quality: Genetic Algorithms and Metamorphic Testing for Software Engineering ML"

Robustness of LLMs for the generation of the algorithm to word with Code2Vec
- gives a good func. name with base variables
- gives nonsensical func. name when a single variable is changed

Parallel to CodeBERT/LaMDA prior work; LLMs are not robust; but imitating human-like changes to redteam them is hard => Purpose of this work.

31/🧵 #GECCO2023 #GECCO

Tests on VGG16 and EfiicientNet; suggest adaptive repair does work better, but hyparparameters have a minimal impact.

30/🧵 #GECCO2023 #GECCO

"Adaptive Search-based Repair of Deep Neural Networks"

NN usage for self-driving vehicles failures as justifying example. Deep Networks repair is the mitigation technique for those problems this paper will be focusing on.

Problem so far: repair is changing the model, and the faults can easily shift.

The idea is to pinpoint the failure at lookin at the neurons differentially activated in failure cases + switch weights of suspicious neurons.

29/🧵 #GECCO2023 #GECCO

Next is "Automated Repair of Unrealizable LTL Specifications Guided by Model Counting".

Reactive systems: Control system primitives; LTL is the language in which goals of the control system are formulated. The idea is to find cases where controller is not implementable, eg due to contradicting goals and/or assumptions.

The goal is to use a GA in order to create a set of requirements that is no longer contradictory.

A couple of convincing benchmarks at the end.

28/🧵 #GECCO2023 #GECCO

Evaluated on the apps on the Android store, crashes found, bigger search budget (generations) works a bit better, no generalizing fitness function (#NoFreeLunch).

27/🧵 #GECCO2023 #GECCO

Time for a yet another change of scenery for a "Search-Based Software Engineering" session, with "Search-Based Test Generation Targeting Non-Functional Quality Attributes of Android Apps".

The goal is to allow to combine the general test properties (crashes), Source code coverage and quality properties (energy usage, memory, compute resources, ...)

STGFA-SMOG Algorithm: basically a variation on the Genetic Algorithm (with crossover and diversification).

26/🧵 #GECCO2023 #GECCO

We are not going to change disciplines where paid journals are already entrenched, but the least we can do is to resist switching to them in disciplines that has been independent from that publication schema.

(oh, and bonus anti-points to Nature for non-anonymous review which further the feudal research and publication schema in the domains relying on paid journals for promotions and funding).

<end rant>

25/🧵 #GECCO2023 #GECCO