Things I’ve learned reading just ONE paper

One of the most important parts of the research job is reading papers. As an early stages of a new research line you are going to deep, first month of reading may seem fruitless, because you can think you are not doing anything, nor developing new experiments or whatever. But, I realized, reading one of my friends’ (and another member of GeNeura) paper how many information you can learn in just one paper. As an example, I am going to write here all concepts I’ve learned reading his paper “EvAg: a scalable peer-to-peer evolutionary algorithm” (you can download it from here).

It is a paper about P2P populations where each individual of a GA is present in a different node (fine-grain) in the network and explain how to interchange information among the nodes. Many experiments are performed and all that stuff, but I am going to concentrate in some concepts, not in the paper results, because Juanlu explained them in previous posts of this blog.

Panmictic population: each member of the population has the same probability to reproduce with the others. Applying to a graph, it is a fully connected graph.

Selectorecombinative GA: Is a mutation-free GA. It is used in population studies, due to all diversity source is from the initial population.

Trap functions: scalable and decomposable functions (divided in Building Blocks, BB). Varying the number k of each block also varies the problem difficult. Fitness depends of the number of 1s in the block. For example, in a k=4. F(0) = 3, F(1)=2, F(11)=1, F(111)=0, F(1111) = 4. It could be deceptive (one of the edges leads to a sub-optimun, and other to a global optimum).

Watts-Strogatts network: It is a small-world graph where nodes forms clusters, like real Internet. It is very easy to create a WS network from a ring topology (just alter each edge to connect another node with a probability p, so some connections are deleted). In these networks, selective pressure is equivalent to panmitic populations, but more scalable due there are a lower number of edges.

Takeover time: time that the best individual take over the entire population.

Metrics for experiments:

  • Success Rate: number of times algorithm finds the optimum.
  • Average Evaluations to find Solution (AES)
  • Best fitness convergence
  • Genotipic entropy: diversity of the population based in the genotipic distance to optimum genome (using Hamming Distance, for example). Wilcoxon test can be used.

As you can see, no information about the results of this paper is presented in this post, I’ve just written about new things I didn’t knew.

Tip: I am using Evernote to take notes of every paper I read. Evernote is like Dropbox, but with notes. I annotate shops, projects, todo lists… and I have full synchronization with all my computers (in Mac OS X, GNU/Linux) and mobile devices (Android or iPhone). There exist many usages of this program, so I recommend to take a look.

Pablo.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s