Between intelligence and evolution

Is there a connection between intelligence and the ability to learn and evolution? And how can artificial intelligence give us an interesting perspective on this relationship?

Israel Benjamin;

At the end of the 19th century, the psychologist James Baldwin (Baldwin) asked what the evolutionary advantage of the ability to learn is: why would such an ability increase the chance of the creatures endowed with it to pass on their genes to the next generation? At first glance, the answer is simple: a creature that avoids repeating errors, and that can repeat strategies that have been proven to be successful, will live longer, will win more competitions for mates, and will therefore have more offspring that will carry its genes - including the genes that contribute to the success of learning. But Baldwin saw that there was a difficulty in this answer.

As we know, poisonous insects exist, and most of these insects have prominent warning colors. In doing so, the insect "informs" the bird: "You shouldn't eat me - I cause a stomach ache!" But how will the bird understand the message? One can imagine a bird whose genes already include a pre-programmed behavior of avoiding eating an insect with warning colors. It is also possible to imagine another bird whose genes allow it to associate a stomach ache with eating a certain insect shortly before. Which planning is more effective? The learning bird will have to go through the agony at least once, unless it is able to recognize suffering in another bird, remember what that bird ate, and connect the things; Or if each bird learns the eating habits of its parents. Learning is also a less reliable process - unexpected events may interfere with the link between cause and effect.

Until the moment of learning, the bird is vulnerable to eating poisons that may cause its death directly or indirectly (for example, by making it easier for a predator to capture it). In contrast, programmed behavior works reliably from the first minute of life. In addition to this, learning requires resources: it may require additional brain tissue, which is expensive in terms of weight and energy consumption. According to these considerations, it can be expected that the "pre-programmed" bird has a better chance of survival, and during evolution its genes will spread in the population at the expense of the genes of the "smart" bird.

Some readers must already be against this argument: there are many things to learn, and some of them may change from generation to generation. The advantage of learning is its flexibility and its ability to deal with a changing reality, even during one generation, compared to the large number of generations required for evolution to take advantage of genetic diversity and mutations in the population to "discover" the genetic code required to program the desired behavior. This objection is not accurate: many things remain constant over the generations, such as the fact that most insects with warning colors are poisonous. Isn't it "payable" for a bird to code such knowledge genetically?

Artificial intelligence, learning and evolution

Before returning to Baldwin's question, we will explain what the artificial intelligence section and issues in evolution are about, by asking a parallel question taken from the world of artificial intelligence: we want to develop software to manage a complicated production machine, which will automatically identify potential danger situations and take the necessary steps to resolve them. One option is to analyze in advance all the options the software might encounter, and write a command in the software to detect and respond to each option. This method requires an omniscient planner (and who has enough time to express all his knowledge in computer commands), because the required response to the situation where we saw both A and B is not the combination of the responses required in situation A only and situation B alone. In other words, the number of options to address may be too large to be practical. The parallel idea in explaining the complexity of animals is creation by an almighty God.

Another option is to write "learning" software. At first she will fail very often, but with feedback on success or failure she will gradually improve. It goes without saying that if we can create a computer simulation of the production machine, we will run the learning software against this simulation and thus save time and damages. One of the most accepted methods for building such software is a "neural network": the software simulates a collection of nerve cells and the connections between them, where some of the neurons are activated by status reports from the production machine, and another part controls the functions of that machine.

The connections between these virtual neurons determine to what extent the activation of a particular neuron will cause the excitation, or suppression, of the action of a neuron connected to it. A network with correct connections will respond correctly to any status report. If the simulation shows that the network reacted incorrectly to a certain situation, the network will "learn" this by changing the pattern of connections between the neurons. If we have well built the network and the processes of updating the connection patterns, then after enough such attempts the network will perform well in most possible situations, even those that did not happen in the learning process.

A third option is to build the software in stages, using a method called a "genetic algorithm": we will start with a collection of programs that differ in their response to each event, and we will examine their performance against the simulation of the production machine. This collection is called "The First Generation". Although none of the programs will be particularly good, we can find the best of them and "hybridize" them by creating a collection of "second generation" programs from them: to build software for the second generation, we choose two "parents" from the more successful programs in the first generation And we will build the behavior of the new software from parts of the behavior of each of the parents, plus some accidental changes ("mutations"). When we examine the performance of the second generation software, we find that some of the offspring are better than their parents and some are less good, but on average this generation will usually be better. We will continue this method for several hundreds or thousands of generations, and it is likely that we will find well-behaved software.

It is clear that the genetic algorithm draws its inspiration from Darwin's theory of natural selection, when better programs get more offspring that are similar to them, and the neural network imitates simplistic models of the human brain and its way of learning.

Interaction between learning and heredity

If it were possible for the learning of one generation to be registered in the genetic load inherited by the next generation, as in the hypothesis proposed by Lamarck, we would be able to gain all the advantages: learning to handle new situations, albeit with limited reliability, along with innate and reliable behavior for situations that the previous generation has already encountered . Darwin still thought that there was also a place for "Lamarckian" evolution, but he believed that the natural selection model had a greater influence on the development of species. After Darwin's days, when the genetic mechanism was understood, it became clear that events during life cannot change the genetic code passed on to offspring.

But even before the discovery of the mechanisms of heredity, Baldwin saw that even in Darwin's theory there is an indirect way in which learning causes evolutionary change. This idea, published in 1896, and proposed separately in the same year by Lloyd Morgan and Osborn, is now called the "Baldwin effect". Consider a population of birds spreading to a new environment, where there is a poisonous insect of a type they have not encountered before. Because of this, the innate behavior should not be expected to protect them: if the birds cannot learn, they may not survive in the new environment.

If they can learn, it is still expected that this insect will cause many of the birds difficulties. These difficulties create an opportunity for evolution: if, in the course of the next generations, a hereditary trait appears in the population that causes the insect not to be eaten, this trait will increase the survival of the birds that carry it, and will gradually spread through the population. In this model, the role of learning is to serve as an immediate, if imperfect, solution to a problem that would otherwise annihilate the group. This solution "buys" enough time for evolution to create a more reliable solution.

Simulation of evolution

The Baldwin effect sounds convincing, but it is not clear how widespread it is, and what part it played in the evolution of species - just because something may happen does not mean that it actually happened. In an attempt to understand the potential power of this effect, some researchers turned to computer simulations. One of the well-known studies was carried out in 1987 by Jeffrey Hinton and Stephen Nowlan (Hinton, Nowlan). They combined a simple neural network, with twenty potential connections between the nerve cells - that is, each such connection can be present (1) or disconnected (0).

The research was structured so that only one option was correct, out of about a million possibilities of combinations of 0 and 1. Each "creature" in the simulation was "born" with a certain structure of connections, but some of the connections were marked as predetermined and some as learnable. During the life of the same creature, it can learn by a simple method: an experiment of up to a thousand different combinations of "existing" or "disconnected" for those connections that were not predetermined. If all twenty connections are learnable, the chance of hitting a XNUMX in a million possibility in a thousand attempts is obviously very small. Having predetermined connections to the appropriate value for the solution will help increase the chance of learning success - for example, if you need to "guess" the correct value only for ten connections that were not predetermined, only one thousand possibilities should be checked. On the other hand, if one of the predetermined relations is set to a wrong value, then surely the study will not be useful.

The experiment began with a thousand different creatures, each with a random link pattern chosen so that 50% of the links were learnable, 25% were set to a true value, and 25% were set to a false value. A genetic algorithm selected those creatures that managed to learn the correct solution in the allotted thousand attempts, and "breeded" them to create a thousand new offspring in each generation. In just fifty generations, the percentage of "correct" pre-determined relationships increased from 25% to about 60%, while the percentage of "incorrect" pre-determined relationships decreased from 25% to zero. In those generations, the average survival increased from a very low value to 90%. In this, the Baldwin effect was observed for the first time in the virtual world - it is the ability to learn that allowed evolution to operate, and one of the results of evolution was the reduction of the freedom to learn by lowering the number of contexts that can be learned.

This initial research was very simplistic in the way that it simulated the mechanisms of learning and inheritance. Later studies, such as that of David Ackley and Michael Littman (Ackley, Littman) in 1992, showed the same effect under more realistic mechanisms. Such simulations help evolutionary biologists explore phenomena and possible explanations under controlled conditions, without waiting for thousands of generations to pass.

Simulation of intelligence

In turn, artificial intelligence people learn from these studies how to write "smarter" software. From the Baldwin effect comes a possible approach to improving software that seeks a solution within a huge space of possibilities - such as the production control software in the example presented at the beginning of the article. The approach is a combination of evolution and learning: instead of testing the performance of the programs in each new generation only once, we will allow the program to learn in a way that is limited in time, and limited by the features it was "born" with, and we will measure the performance of the programs in each generation at the end of the learning process. It is this performance, after the self-improvement phase, that will determine which of the programs will contribute parts of themselves to the programs in the next generation.

Such mixed entities have become quite popular in recent years. In one example of many, this year a group of students from Bar-Ilan University used a similar idea to solve a complicated problem in the field of travel route optimization, under the guidance of the author of this article. The three students - Liat Barzilai, Reuven Dagan and Uzi Zahavi - tested an innovative method of optimization. Without the addition of the self-improvement step, the best solution obtained by this approach was significantly inferior to the best solution obtained by previous methods. With self-improvement of each solution before moving to the next generation, a "new record" was eventually reached in the quality of the solution, in an incredibly short calculation time.

The development of natural intelligence

The story about the birds and the poisonous insects seems to lead to a contradiction between pre-programming and learning: dangers can be avoided either by genes that dictate innate behavior, or by genes that build nervous systems that are capable of learning, and hence there is no need for both. The Baldwin effect stands in stark contrast to this apparent contradiction, and even explains how those nervous systems, which are extremely complicated in their genetic coding, can develop.

According to Steven Pinker, in his book "How the Brain Works" (1997): "The Baldwin effect probably played an important role in the development of the brain. Contrary to the common assumption in the social sciences, learning is not a peak of evolution achieved only recently by humans. All animals, except the simplest, learn... If the ability to learn had existed in the ancestor of the multicellular animals, it could have guided the evolution of the nervous systems into their specialized circuits, even when these circuits were so complex that natural selection would not have found them by itself." Hinton and Nolan's experiment demonstrates this idea by evolving a neural network with correct links while utilizing a very simple ability - random trial and error.

And what if the neural circuits that develop in this way are themselves used for learning? Like Baron Münchausen extricating himself from the mud by grabbing his hair and pulling hard upwards, we may owe our prodigious learning ability to the primitive learning of a very ancient ancestor. Perhaps it is possible to imitate such an evolutionary-learning process to create "real" artificial intelligence.