AI Glossary / Deep learning / Advanced
Promoter based genetic algorithm
Definition
The promoter based genetic algorithm, or PBGA, is a neuroevolution method that uses promoter-like switches to control which encoded neural units are expressed in a candidate network. Inactive units can remain in the genetic representation even when they are absent from the evaluated network.
What it means
A genetic algorithm searches through a population of encoded candidates. In PBGA, the stored encoding and the active neural network are deliberately different objects. The genotype can retain functional blocks that the current phenotype does not express.
Bellas and Duro developed the approach in work on neural models for changing environments. Later research with Becerra examined promoters and functional introns for non-stationary learning. The purpose is to help preserve potentially useful structures instead of requiring every encoded unit to contribute to the current task.
The biological vocabulary is an analogy used in a computational encoding. These are software representations, not biological genes. Nor is PBGA the same as every genetic algorithm that optimizes a neural network: promoter-controlled expression is its defining feature.
How it works
In the described encoding, a functional block represents a neuron together with incoming connections and associated parameters. A promoter field controls whether that block is expressed when the genotype is translated into a network.
Candidates are evaluated through their expressed networks. Selection, crossover and mutation generate subsequent candidates, with different treatment possible for promoter fields and numerical parameters. Changing a promoter can alter a whole functional unit rather than just one weight.
An inactive block may retain information that becomes useful when the objective changes. Its current effect on the evaluated network is suppressed, but this does not make it immune to all evolutionary changes or guarantee its later usefulness.
The approach has been used within the Multilevel Darwinist Brain research architecture. That larger architecture also has memory mechanisms of its own. Genotypic retention inside PBGA should not be confused with an external store of previously successful complete models.
Three practical examples
- A research experiment compares adaptation when a target function alternates between previously encountered regimes.
- A robot-modeling study evolves networks that predict observations from interaction data in a changing environment.
- A teaching implementation examines how active network size can change while inactive functional blocks remain encoded.
Think of it this way
Think of a workshop keeping several reusable tools in storage while placing only the tools needed for today's job on the bench. Changing the selection can make an old tool available without rebuilding it from nothing.
The analogy captures retention versus current use. PBGA does not know that a stored block is valuable in the way a craftsperson knows a tool's purpose. Evolutionary operators, current fitness and the surrounding encoding determine whether a block survives and whether reactivating it helps.
Worked example
Stored blocks and expressed units are different counts
Consider an invented chromosome containing three candidate hidden-unit blocks, A, B and C, plus a fixed output mechanism. For illustration, use promoter states (1, 0, 1), where 1 means expressed and 0 means inactive. The genotype stores three blocks, while the active network uses A and C.
Suppose a trial changes the promoter states to (0, 1, 1) without changing the stored numerical parameters. The active network now uses B and C. A has become inactive, but its encoded parameters have not been deleted by this stipulated promoter change.
Both phenotypes contain two active units, yet they need not compute the same function. If a new task favors the second network, selection may prefer it. This example demonstrates expression control only; it does not reproduce the complete PBGA encoding, mutation schedule or experimental results. A retained block must still be evaluated when reintroduced.
Illustrative case study
Fictional example: Fern Robotics studies a simulated sensor model whose target relationship changes between two operating regimes. Its engineers compare an ordinary evolutionary baseline with a promoter-controlled encoding under the same evaluation budget.
The promoter-based candidates can keep inactive neural blocks, so the team tests whether returning to an earlier regime requires fewer evaluations to recover useful predictions. It also checks the error immediately after a switch, the population's diversity and the storage and decoding costs.
One retained block performs poorly when reactivated because the surrounding active network has changed. The team therefore avoids treating inactive code as a guaranteed backup model. It distinguishes retention of a component from successful restoration of a complete behavior.
The study's conclusion would depend on the measured comparisons. The presence of promoters alone cannot establish superior adaptation or eliminate forgetting.
Common misconceptions and mistakes
- An inactive unit has been erased. It may remain encoded while being excluded from the current phenotype.
- Retained genetic material is a guaranteed memory. Survival, re-expression and usefulness are separate questions.
- PBGA is simply dropout. Dropout is a different training mechanism; PBGA evolves expression within a genetic representation.
- Biological terminology proves biological realism. The computational mechanism uses selected ideas, not a complete model of gene regulation.
Business use
PBGA is a specialized research concept relevant to adaptive modeling and evolutionary architecture search. Its business significance lies in the broader question of whether reusable computational structure can be retained as conditions change.
Any practical proposal should compare adaptation quality, evaluation cost and maintenance complexity with simpler alternatives, including retraining or explicit model storage. Use non-stationary test sequences that reflect the intended changes rather than only one static benchmark. State whether an apparent benefit comes from promoter encoding, an external memory mechanism or another part of the surrounding architecture.
People also ask
Does PBGA evolve only connection weights?
No. Promoter-controlled expression can also change the active neural structure.
Does a disabled block contribute to the current prediction?
In the intended expression mechanism it is excluded from the expressed network, although exact decoding rules belong to the implementation.
Does PBGA automatically identify every recurring task?
No. Its retention mechanism does not provide a universal task-discovery or adaptation guarantee.
Check your understanding
Two candidates store the same three blocks but express different pairs. Must they have the same predictions because their active unit counts match?
Answer: No. The identities, parameters and connections of the expressed units matter. Active size alone does not determine the network's function.
Related terms
- Genetic algorithms
- Mutation
- Neural Architecture Search
- Continual Learning
- Neuroevolution
- genotype
- phenotype
- genetic algorithm
- non-stationary learning
Learn next
Read genotype and phenotype in evolutionary computation. Draw a chromosome with active and inactive neural blocks, then distinguish changing a promoter, changing a numerical parameter and deleting a block.
More in Deep learning
- Pooling Layer - A pooling layer summarizes groups of values in a neural-network representation.
- Preprocessing - Preprocessing is the preparation and transformation of data before it is used by a model or another analysis step.
- Probabilistic Topic Model - A probabilistic topic model is a statistical model that represents a collection of documents through latent themes and probability distributions.
- Proportional Hazards Model - A proportional hazards model describes how characteristics multiply the instantaneous rate of an event among cases that have not yet experienced it.
- Pulse-coupled networks - Pulse-coupled networks are networks of interacting units that influence one another through pulses.
- Quantum neural network - A quantum neural network is a learning model that uses quantum operations in a network-like or trainable computational structure.
Sources
Bellas and Duro, Modelling the World with Statistically Neutral PBGAs: https://www.gii.udc.es/img/gii/files/bellas02PBGAb.pdf. Bellas, Becerra and Duro, Using promoters and functional introns in genetic algorithms for neuroevolutionary learning in non-stationary problems: https://www.sciencedirect.com/science/article/abs/pii/S0925231208005481. GII research context: https://giilab.eu/en/subline/mecanismos-cognitivos/. Block counts and the robotics case are invented illustrations.
Previous: Projective Rule Next: Prompt