Symbolic Regression and Multi-Gene Genetic Programming

I'm back with my blog being about something useful. Sorry for the long waiting, I was really really busy. Today's topic will be about a class of techniques for speeding up Genetic Programming algorithms in Symbolic Regression scenarios.

Regression vs. Classification

First of all, let me do a quick introduction/clarification regarding regression and classification. These two are machine learning tasks, similar in part, but different in the end. In both tasks we have a bunch of objects each described by a set of numbers (not necessarily but for the sake of simplicity numbers will do great).

In classification the goal is to assign a class to an object based on the numbers it is described by. The class can be any discrete value. Examples are: true/false, red/green/blue, high/middle/low, 1/2/3/4/5/6/7/8/9/10 rating, or any label. The special case with just two classes also has a special name binary classification. Typical classification problem is spam detection - is an e-mail spam or ham?

In regression the goal is to assign a numerical value to each object, again based on the numbers describing that object. Generally, these assigned numbers can have any value. Typical regression problems are various prediction and estimation problems. For example, suppose there is some complicated machine and you want to predict fuel consumption based on some operational data.

Symbolic Regression

Symbolic Regression is a class of tasks for GP systems. Actually, it is the very first example I mentioned in Evolution in a Computer, pt. 3 - we have a bunch of data points and we want to find a function that fits them accurately (enough). It doesn't matter where the data points come from, be it anything from a demonstrative problem as in that blog post to data from medical scans where we need to asses a patient's condition. Whenever we want to find a function fitting some data, it is symbolic regression.

The word symbolic is quite important. It means that the output model is a mathematical formula, that could be, in principle, read and interpreted by a human, or, for example, be given to a computer algebra system (CAS) to compute derivatives, find roots, etc. This is in contrast with, e.g., neural networks which is characterized by its topology (how many neurons are there and how are they connected) and the values of the weights (the degrees of how pairs of neurons are connected). That is not something one could look at and see a pattern nor could be given to some CAS even though it still does regression.

So, from now on, let's talk about Symbolic Regression, or SR in short.

Problem of SR

If you played around with the SR example from the already mentioned previous blog post, you might have noticed that it does not work very well. In the better case it just took quite a long time to get to some sensible function. In the worse case it failed completely and got stuck with some obscure function.

There are several reasons why this does not work that well. I would like to write about one of them and that is fine-tuning the parameters. Vanilla GP works with a set of terminals and non-temrinals and evolves functions by juggling around with pieces of mathematical expressions to hopefully get meaningful results. However, this juggling is often also the only mechnism of how to get some specific numbers in the final function. Imagine, that the true function (i.e. what we would like to find by evolution) is \( f(x) = \sin(3.18x - 6.57) \). However, the GP algorithm is very unlikely to have the 3.18 and 6.57 numbers in its terminal set. It must mainpulate what it has to get (close to) these numbers.

Multi-Gene Gentic Programming

The idea of Multi-Gene Genetic Programming, or MGGP for short, is to decompose the problem into more, presumably simpler, subproblems. However, the decomposition is not exact, as (almost) nothing around GP is. MGGP addreses part of the problem with finte-tuning parameters.

How MGGP works

In classical GP, each individual is one tree which is also the solution to the problem, in case of SR it is the expression which is supposed to describe the data and its output is the regression estimate. In MGGP, in contrast to GP, each individual is made of multiple trees which together make the solution. These trees are called genes.

But how do we get a single number out of multiple expressions? The answer is linear combiantion. That means that the output of every gene is multiplied by some constant number and then they are all summed up together. It is basically a weighted sum of the genes. More, in MGGP, there is always an implicit gene which is always equal to the number 1. The reason for such gene is that we get an additive factor for free.

Example: let's have these three genes and (for simplicity) there is only one variable to work with, \(x\): $$ g_1(x) = x $$ $$ g_2(x) = \sin(x) $$ $$ g_3(x) = \cos(x) $$ and of course there is always the implicit gene of 1, let's give it the index of 0: \(g_0(x) = 1\). Now, to get the final expression, there are some constants, let's call them \( w_i \) which multiply the individual genes. The final expression would be, in this case: $$ f(x) = w_0 g_0(x) + w_1 g_1(x) + w_2 g_2(x) + w_3 g_3(x) $$ or to make it simpler (because \( g_0(x) \) is always going to be 1): $$ f(x) = w_0 + w_1 g_1(x) + w_2 g_2(x) + w_3 g_3(x) $$

By now, you should have an idea but you should be wandering: Where do we get the constants (\( w_i \)) from? Yes, that is the key for MGGP. And the trick is really simple. We use the power of linear regression. Linear regression is a kind of regression analysis (see, SR is another kind of regression analysis) which produces models in the form of linear combination of the input features. So, for example, if our problem has two input features, the linear regression produces models like \( f(x_1, x_2) = w_0 + w_1 x_1 + w_2 x_2 \).

There are several ways of how to get the constants \( w_i \) but probably the most widely used is the ordinary least-squares solution. Let the input features be arranged in a matrix \( \mathbf{X} \) so that each datapoint (its input features) is one row of the matrix and let the target values be arranged in a column vectory \( \mathbf{y} \) so that \(i\)-th element corresponds to the \(i\)-th row of the matrix \( \mathbf{X} \). If we have these, the vector of the constants, also called weights, can be computed as:

$$ \mathbf{w} = (\mathbf{X}^T \mathbf{X})^{-1} \mathbf{X}^T \mathbf{y} $$

This way we get the weights that multiply the input features but we don't get the additive factor, \( w_0 \). This is fixed easilly by prepending a column of ones to the matrix \( \mathbf{X} \).

Identical procedure is applied in MGGP. Instead of using the vanilla input variables, we first compute “new” features by computing the outputs of the genes - each gene produces one new feature - and of course create the implicit feature of 1. Important fact is that the computation of the weights is done at evaluation, i.e. they are always computed optimally with respect to the genes that are available in the individual.

Additional operators

Since there are multiple genes, the crossover operator is split into two types:

• Low-level crossover - this is a classical crossover known from GP: a random node is chosen in each individual and the subtrees rooted in these nodes are swapped.
• High-level crossover - a new type of crossover arising from the fact that there are multiple genes: whole genes (one or more) are swapped between the individuals.

What is it good for

The major benefit of this approach is that it lifts the duty of finding the linear parts from the shoulders of the underlying GP which needs to focus on the non-linear parts only. Also, it makes a kind of decomposition of the problem - multiple, maybe smaller, genes work together to create a reasonably good solution. It addresses the problem of fine-tuning the parameters to a ceratin degree - the top-level parameters, i.e. the linear coefficients, are not needed to be found by evolution. However, the constants hidden inside the non-linear parts are not adressed by this mechanism and are sought traditionally (in GP terms).

References

• GPTIPS - the leading implementation (and maybe the only one really used) of (not just) MGGP by Dominic Searson.
• Hinchliffe MP, Willis MJ, Hiden H, Tham MT, McKay B & Barton, GW. Modelling chemical process systems using a multi-gene genetic programming algorithm. In Genetic Programming: Proceedings of the First Annual Conference (late breaking papers), 56-65. The MIT Press, USA, 1996.
• Searson DP, Willis MJ & Montague GA. Co-evolution of non-linear PLS model components, Journal of Chemometrics, 2, 592-603, 2007.

Quickie: zegkljan.net

Hi folks, this is just a quick one. I have finally set up my own domain zegkljan.net. So, from now on, you can find my blog at blog.zegkljan.net. The domain itself currently redirects to the blog but in the future I want to have some other stuff there as well. Of course, the original address zegkljan.blogspot.com still works and redirects to the blog.

Evolution in a Computer, pt. 3

After months of waiting, I welcome you to the final part of my humble introductory mini-series about evolutionary computation. And today's topic is Genetic Programming

Discovering functions

Recall the example of finding the minimum of the function from the first part of this mini-series? We had a function we didn't know anything about (actually, we did, but that was just for the sake of showing what we can do when we know) and we wanted to find a minimum of that function. Now, let's look at a different problem: we don't have the function at all, only a handful of \(x, y\) pairs, and we want to find out what the function actually looks like. How to do it?

If we restrict ourselves to a specific set of functions, e.g. linear functions, we can use specialized methods like (multiple) linear regression that will just throw the function at us given the data. Period. However, this has the problem of restricting to only a specific set of functions. If the "secret" function was actually a linear function, it would be the best we could do but if it wasn't, we would get errorneous function.

We could take a different approach and use more complicated machine learning methods like Artificial neural networks (ANN) that have the universal approximation property (simplified: given enough neurons they are, in principle, capable of representing almost any function with arbitrarily low error). However, their problem is that they can be immensely complex and if one wanted to use it as a mathematical expression, it would be mess.

As you might already have guessed, we are going to use.. Evolutionary algoritm! Well, we know how to evolve binary strings and real numbers... how do we use it to find a function? Well, it is very complicated. You could design some algorithm that would translate binary string to a function but it would probably be very messy and not efficient, not speaking about the fixed length of the binary strings which would be a major problem. However, there is a representation that can be used... trees!

Tree representation

A mathematical expression can be described like a tree. Let's take some made-up expression like $$ (x + \sin{y})^2 - \cos{3x} $$ Where is the tree in there? It's easy - the operators (i.e. addition, mutiplication...) and functions (i.e. sin, cos...) are inner nodes and the variables (i.e. x, y...) and constants (i.e. numbers) are leaves of the tree. The corresponding tree for the expression above looks like this:

Evolution of trees

So now that we have a representation for math expressions, how do we actually evolve it? Well, pretty much in the same way as in classical GA - by selection, recombination and mutation. Selection is identical - the better solutions (in this case those that have lower error) are preferred more than the worse ones. Recombination, i.e. crossover, and muation are the more interesting ones.

Crossover

There are more types of crossover but I'm going to speak about the most basic type - the subtree crossover. This crossover works in the following way:

1. Pick a random node \( N_1 \) in the first parent
2. Pick a random node \( N_2 \) in the second parent
3. Exchange the subtrees rooting in \( N_1 \) and \( N_2 \) between the two parents

Look at this example: And that's it! It's quite simple. Sometimes some tweaks are used, e.g. giving more preference to nodes that are deeper in the tree, but in essence, the basics are the same - just swapping subtrees.

Mutation

As there is subtree crossover, there is subtree mutation too. Again, it is based around mangling subtrees. However, contrary to the crossover, in mutation we don't exchange subtrees between multiple solutions but we modify only one. The most basic subtree mutation looks like this:

1. Randomly pick a node
2. Throw away that node (and its subtree) and replace it with a new randomly generated subtree

No illustration here - it's really that simple.

Demo

This is a little demo of GP in action. The task is to find the function based on a bunch of samples (the red points; the sought function is the dashed curve). The actual function that we want the GP to find is $$ f(x) = 5 \sin{x^2} $$ ATTENTION: you might find that the algorithm "runs away" with bigger and bigger and messier and messier functions. I suggest you stop the algorithm and start again. This algorithm is really not tuned very well, it's just a demo.

Clicking Start starts the evolution. Clicking Cancel during the evolution terminates it. You can play with the population size (must be positive even number), crossover and mutation probs (must be between 0 and 1) and the tournament size (must be between 2 and size of population). The evolution runs indefinitely (i.e. until you stop it). You can also play with the number of available data points (the more the easier it should be for the algorithm to find the solution). This thing is written in Dart and you can find the source here.

Not just functions

We can evolve functions. But one could say “Fine, but where's the programming in that?” and she'll be totally right. So far we can make mathematical expressions. But the tree representation goes beyond that. All you need to do is to represent a program like a tree. And it is quite simple. Let's imagine that we control an ant in a 2D grid world. The ant can turn left, turn right and decide what to do based on whether there is a piece of food just ahead of it. And here is what a program in this domain might look like: As you can see, the only difference is that the branching point (or non-terminal) is not an operator but a decision-making point and the leaves are not variables and constants but actions. In fact, there could be many more non-terminals, e.g. cycles, function calls etc.

Closure property

All the Genetic Programming I introduced above is called Koza-style GP because it is the idea of John R. Koza which he published in his book [1]. For this type of GP to work a so called closure property must hold. Closure property holds if an output of every terminal and non-terminal is acceptable by every non-terminal as an input. That menas that whatever our nodes output, it must be possible to chain it with an arbitrary node.

If our non-terminal set is made of plus, minus and multiplication and our terminal set of variables and numbers, closure property holds - however we add, subtract or multiply numbers, we always get a number that can be furter added to, subtracted from or multiplied. However, if we extended our non-terminal set by a non-terminal such as if that would take three children - a condition, a result when condition is true and a result when condition is false - then the closure property does not hold - the condition requires a boolean value but that is different from numbers. We would have to employ some rule that would define true and false in terms of numbers.

In fact, it would be enough to just add division to the non-terminal set. Yes, it would still produce only numbers but what if the second argument to a division resulted to zero? We can't divide by zero so division cannot accept zero as its second argument. You see, it is sufficient that a non-terminal doesn't accept some of the values even though they are of the same type.

Summary

In this long-awaited (I hope :)) post I described a little bit of what is known as Genetic Programming. It is basically a way of evolving mathematical expressions, computer programs or basically any tree-like structure using the principles of genetic algorithms. It might (or might not) seem as a brilliant technique able to solve anything... but don't let yourself be fooled. Finding functions and programs is really hard and most of the time this vanilla GP doesn't work very well. But it's a start and it can solve at least some easy problems.

Regarding further posts I'm going to publish here, I don't have any clear plan. I suppose I might dive into specific algorithms or things I do or something totally different. We'll see :).

---

^[1] Koza, John R. Genetic programming: on the programming of computers by means of natural selection. Vol. 1. MIT press, 1992.

Evolution in a Computer, Pt. 2

After a long waiting, it's here! To seamlessly continue from the previous part, I prepared a little demo of a genetic algorithm. What we have here is a so-called Facility location problem, or a variation of it.

There is a bunch of cities (50 in our case) and we want to supply the cities with goods so we must open some factories. We must decide how many factories we are going to open and in which cities. We won't explicitly assign which cities are to be supplied by which factories but a city will be assigned to the nearest factory (imagine the factories have unlmitted capacity). The catch: we want to minimize the total cost:

• every opened factory costs us some fixed amount of money
• delivering goods from a factory to a city costs us an amount of money equal to the (euclidean) distance from the factory to the city

Clearly, we cannot try every combination of cities with a factory as it's \(2^{50} \approx 1.13 \cdot 10^{15}\) combinations. That would took almost 36 years if checking one solution took one microsecond. Yes, this problem is NP-hard. Instead, we are going to use a... Genetic Algorithm!

So how do we represent a solution to our problem? Well, we'll use a binary string, as in proper GA, with the number of bits equal to the number of cities. Then \(1\) in position \(n\) means that the \(n\)-th city has a factory and vice versa, e.g. a string like 001000110 would mean that the third, seventh and eighth city would have a factory while the first, second, fourth, fifth, sixth and ninth wouldn't. Now, give it a try!

Clicking Start starts the genetic algorithm from beginning with the current map. Clicking New map generates a new (random) map. Clicking Stop during the optimization stops it. You can edit the cost of opening a facility, size of population (must be positive even number), probability of crossover and mutation (must be between 0 and 1) and size of tournament (must be between 2 and size of population). The evolution runs indefinitely (i.e. until you stop it). The world is 100 by 50 units. This thing is written in Dart and you can find the source here.

Real-valued Evolution

So far we have spoken about a genetic algorithm that uses binary strings. This means that our solutions are inherently discrete. But what if we wanted to really find the minimum of the function from the part one?

The „standard“ way

The first way of dealing with real number is to just express them using a binary standard, e.g. the IEEE 754 single-precision floating-point format. But this has problems. Imagine numbers 1.763 and -2.611. These would look like this (e.g. using this tool):

 1.763 = 00111111111000011010100111111100
-2.611 = 11000000001001110001101010100000

And now imagine what would happen when we crossed them over using single-point crossover after the 5th bit:

00111111111000011010100111111100
11000000001001110001101010100000
turns into
00111000001001110001101010100000 = 3.9840699e-05
11000111111000011010100111111100 = -115539.96875

One could say "Well, that escalated quickly!" and indeed, these problems arise all the time. If we explore all the possible crossovers (there are 32 of them resulting in 64 numbers) they include gems like -3.252161001305617e+19, 3.591135600494495e-39 and NaN. Bit-flip mutation is going to have similar problems.

The good way

Much better way than sticking to the binary representation is just use the real numbers. Then, if we have a one-dimensional problem (like in searching for the minimum of that function), the genotype is just one real number. This brings a lot of changes in the crossover and mutation operators.

As we have seen, just crossing over the binary representation is bad. Therefore we need to find a better way of combining two individuals. We would like the offspring(s) to be „close“ to the parents, but be different from them. Since we are in the real numbers, we can use the power of probability and random sampling! One of the simplest real-valued crossover is a so-called arithmetic crossover, which is basically a weighted mean of two (or more!) parents with randomly chosen weights. Or more formally:

$$ \mathbf{x}^{Offs} = \sum_{i=1}^{N_{Pa}} \alpha_i \mathbf{x}_i $$ where \(\sum_{i=1}^{N_{Pa}} \alpha_i = 1, \alpha_i \in (0, 1)\); \(\alpha_i\) are picked randomly for each offspring.

And there are many other crossover operators, e.g. BLX-α, SBX, PCX...

The mutation is easier. One of the most used way is to just add a randomly generated number (or vector if our genotype is more numbers) from a distribution of our choice. The most used distributions are the uniform distribution (where we need to specify the bounds) and the normal distribution (where we need to specify the variance).

Next...

The final part of this introductory mini-series, which will finally get us to what I'm actually doing, is going to be big. It is going to be about Genetic Programming. So stay tuned!

GECCO - days 3, 4 and 5

I'm very sorry I didn't post updates on GECCO as it went. I was quite busy at the conference as well as tired at the end of the day(s). So I'll sum up the three days of the main conference in this one post.

Day 3

It was the first day of the main conference, i.e. the part where the full papers were presented in a 25 minutes long talks (20 + 5 for questions and discussion). It started with a keynote by Ricard Solé from Universitat Pompeu Fabra entitled „Re-designing Nature with Synthetic Biology: from Artificial Ants and Tissues to a New Biosphere“. He talked about designing new organisms to do useful stuff like fixing cracks in concrete or repairing damaged land. To be honest I didn't understand most of it very much but it was interesting.

Here are two interesting talks I have visited. They are not the only ones but these I find the most interesting and having the highest impact on my own research.

Building Predictive Models via Feature Synthesis

A great talk given by Una-May O'Reilly, from MIT, introduced a new approach for building predictive models. It is based on Genetic Programming and LASSO, a linear regression method that penalizes coeffitients size. I still have to read the paper but from what was told during the talk, but this is really a step forward in evolutionary machine learning as it is supposed to be fast and produce good models.

Examining the „Best of Both Worlds“ of GE

This talk I would call „a bad day for GE“. GE, or Grammatical Evolution, is an algorithm for Genetic Programming which happens to be the one I use very extensively (I promise I will write about this! Really! Sometimes...). In his talk, Peter A. Whigham has shown that the GE is not as good algorithm as it is thought of. In fact he has shown that it is only a little bit better than random search. If one wants to use grammar-based GP approach he recommends to use Context-free Grammar Genetic Programming (CFGGP), an approach older than GE.

Day 4

The fourth day, or the second day of the main conference, started with a keynote by Kate Smith-Miles from the Monash University, entitled „Visualising the Diversity of Benchmark Instances and Generating New Test Instances to Elicit Insights into Algorithm Performance“. The title sounds scary but it really is not that. Kate described the problem of current algorithm research procedure:

1. Propose a new algorithm which is (usually) a minor variation on the existing ones.
2. Measure the performance on this, this ... and this benchmark.
3. Report that my new algorithm is overall a little bit better than the state of the art.
4. Publish a paper.

She argued that this approach can only show strengths of the algorithm and not its weaknesses, and she raised a question whether the benchmarks that are commonly being used are really that diverse. She continued on with her talk by demonstrating a new approach for performance measurement of algorithms by computing features of the problems, projecting this space into 2D and there visualising the areas where particular algorithms are the most effective and where lay their weaknesses.

Interactively Evolving Compositional Sound Synthesis Networks

An interesting talk at the Digital Entertainment Technologies and Arts track given by Amy Hoover. The research she presented was about using CPPNs evolved by NEAT to produce sound synthesisers. All this happens through an interactive evolution, i.e. the user manually selects which „solutions“ are good. After all, you can try the very product of the research here. http://bthj.is/breedesizer/

An Efficient Structural Diversity Technique for Genetic Programming

The first of the three best paper nominees in the Genetic Programming track. The talk itself was not that great but the presented research, or the idea behind it, is brilliant. It was about enhancing diversity in GP, but with a very very simple yet powerful mechanism: you record first (few) levels of the trees, called genetic markers, and you record a „density“, or how many individuals have the same genetic marker, and you then use some multi-objective paradigm to optimize both the fitnes and the density.

Genetic Programming with Epigenetic Local Search

Second of the best paper nominees. The talk was about an epigenetic search in GP. In biology, epigenetics is a subfield of genetics which looks at the changes in gene expression other than the changes in the DNA (very simplified). Here it meant the ability of parts of the genotype to be simply switched off. Since this would be very difficult to do with classical tree GP, the Lee Spector's Push language was used. The genotypes simply carry a bit for each element in the genotype which simply means whether this element is to be used or not. The epigenetic local search then plays with these bits.

Memetic Semantic Genetic Programming

The last of the best paper nominees and actually the one that won (though I voted for the first one). The research was focused on GP for boolean algebra (and extended for finite algebras). The „trick“ was to analyze where an individual is wrong and repairing the subtrees that cause it to be wrong, using a static library of trees generated once at the start.

Day 5

The last day of the conference. It started with a „ceremony“ where best paper winners were presented, as well as the winners of competitions and other awards. Then the next GECCO place was presented, which is going to be Denver, Colorado, USA.

After the ceremony there was the last keynote, given by Manuel Martín-Loeches from Complutense University of Madrid entitled „Origins and Evolution of Human Language: Saltation or Gradualism?“. And as the title says, it was about the evolution of the human language. It was very interesting and the message is that Manurel Martín-Loeches stands by the theory that the language was a result of gradual evolution by small mutations, rather than one (or a few) big macro-mutation(s).

The last session of GECCO did not have any talks that were very relevant for my research so I looked mostly for things that were just interesting...

Novelty Search for Soft Robotic Space Exploration

An interesting talk about using Novelty Search for evolving soft robots for space exploration. The research was focused on evolving robots made of soft materials of varying properties for exploration of other bodies, like the Moon or Mars. It used indirect encoding by means of evolving CPPNs that produce the robot instead of evolving the robot directly. The CPPN told whether a voxel is to be filled with the material and if yes then with which material. The presentation was quite funny because the presenter showed us videos of the evolved robots moving, some of which being very funny.

Final summary

This was my first GECCO and also my first conference of such magnitude. In my opinion the conference ran very smoothly, there were no problems. I successfuly presented both my papers, one at the Student Workshop and the second one in the regular poster session. I had a chance to talk with other researchers and students which was quite nice. To sum up, I consider GECCO 2015 to have been great and I hope it was not the last one for me.

GECCO - day 2

„Hola!“ as they say here in Madrid. The day two of GECCO 2015 is over. Well, still not the GECCO itself but the workshops and tutorials before. I presented my student poster today and I would like to mention two of the tutorials I attended.

Evolutionary Computation: A Unified Approach

Given by Kenneth A. De Jong, it was a nice and simple tutorial about the general aspects of evolutionary computation. The main goal of the tutorial was to give a broad overview of various characteristics and aspets of evolutionary algorithm, evolution strategies, genetic algorithms etc. and extract all of these into a „unyfing“ field called Evolutionary Computation. The tutorial was very basic and I can't say I was taught something new, but Kenneth's overall presentation put all the things into a single perspective which was very good and excellent to sort thoughts out.

Expressive Genetic Programming

Given by Lee Spector (GitHub), this was also a nice tutorial heavily targeted at the Push programming language designed for genetic programming. The language is, one could say, weird: everything is on stacks, including the program itself, everything reads from stacks and pushes to them or somehow mangles with them. It is designed with minimal syntax constraints. It is for a reason - so that random programs can exist and be valid. It's worth giving a look. There are many implementations of Push, like Clojush which is in, you guessed it, Clojure! I decided I'll try to implement it in Haskell and name it Hush :). I already created a GitHub repo, but there is nothing there yet.

Stay tuned for the report from today which is the third day and also the first day of the main conference. It is (almost) filled with interesting talks, unfortunately some of them I will have to miss (because there are other ones taking place at the same time slot). Stay tuned...

Hello from Madrid! GECCO - day 1

Hello everyone from Madrid, Spain! First of all, it is quite hellish weather here - no clouds whatsoever and close to 40°C (I won't tell you Farenheits because I hate that!) in the shadow. At least there is some wind blowing... Anyway, today was the first day of GECCO, or more precisely the workshops and tutorials before the main conference. My main focus was the Student Workshop where I presented my paper which I won't write about yet because the Evolution in Computer, pt. 3 is not out yet. As a matter of fact, pt. 2 is not out yet too for which I'm very sorry (but it's almost ready to be released, maybe I'll post it while still in Madrid, stay tuned).

Low or No Cost Distributed Evolutionary Computation

One of the very few tutorials I had time to attend was this one given by JJ Merelo of Universidad de Granada.

The @gecco2015 talk on "low or no cost distributed evolutionary algorithms" http://t.co/ixkMNaeeRs (+ last minute changes) /cc @jormaral

— JJ Merelo (@jjmerelo) July 11, 2015

It was about how to use (almost) free (to you) resources to do your evolutionary computation job. He claimed to use JavaScript because it is a standard, it is everywhere and you can write both client and server app (through node.js) and that you should to it in JavaScript because in the end you must write both the client and server so why not do both in the same language. Good point, but he probably does not know about Dart. That one has all the attributes of JavaScript he presented and it is not **** :). By the way, if you want to know more about Dart and things around, look at a blog of a friend of mine, Jana Moudrá, who is also a Google Developer Expert on Dart.

Back to the tutorial. The main point was to distribute the evolutionary computation using pool EA. Unfortunately, he didn't talk about that, instead he talked about the technical realization of such distributed computation. The idea is to use web browsers, which are the new operating systems, of ordinary people to run parts of your computation. In other words, turn people into computing power by „convincing“ them to visit your page that contains the client code that will make them take part in the distributed computation. The second main idea is to use (almost) free services to run the server part (he mentioned Heroku and OpenShift). The third idea is to tell about this to people using social networks (he mentioned Twitter) to get as many „clicks“ as possible. But not to smuggle the dirty computation hidden somewhere in an otherwise innocent page, but be open, honest and give „something“ in return. That something is to be knowledge, e.g. code on github.

Should I rate the tutorial, I would say I'm a little bit disappointed. I expected more about the algorithmic part, but on the other hand the idea of using web browsers of ordinary people is brilliant.

That concludes the day one at the Genetic and Evolutionary Computation Conference or GECCO in Madrid, Spain. Stay tuned for the next day.