Summary

It uses neuroevolution – a genetic algorithm to evolve the architecture of a neural network – to optimize the decision-making of a rat inside an elevated plus-maze.

Notes

  • It uses a Genetic Algorithm to evolve not only weights but the architecture of a neural network. Mentions another paper (Floreano, Dürr, & Mattiussi, 2008) saying that this process is called “Neuroevolution.”
  • Used the same methods – including the fitness function – from authors’ previous two publications. Only the constants are a little different: αo=0.015, αe=0.011, αc=0.012, γ=2 and β=2.
  • Elevated plus-maze is divided with five squares in each arm. This work uses a reduced maze, with only one enclosed arm, and another open. This comes from the consideration that the two equal arms can be considered as a single one.
  • The neural network is an Elman Network, with one hidden layer. The outputs of the hidden layer are reinserted as inputs. Output has four neurons: standing, rotating 90º to the right, rotating 90º to the left, or going to the next square in the current direction. Six inputs, being three short-distance (10cm) and other three long-distance wall sensors (40cm ~ 4 positions from the rat).
  • Minimum of one sensory input, and hidden neurons, is one. Weights are real numbers between -5 and 5.
  • Population of 500 individuals. Each has two chromosomes: one is a binary array with ten positions, indicating which of the sensors are active; the second contains the weights. Evolved for 15,000 generations, with each rat staying in the maze for 300 time steps or 5 minutes. Two individuals selected by etilism, other two by tournament. The one with the best fitness has 75% of chances to be selected. One-point crossover with 0.6 of probability. Flip mutation in the first chromosome. Gaussian mutation with zero mean and unit variance applied to the second, with 0.05 probability.
  • Average of 30 runs was compared to a group of 27 rats. Used Mann-Whitney test with significance level 0.05 to do the comparison.
  • Considered as the most important characteristics for comparison: number of entries and time spent in each of the arms.
  • When compared to real rats and model using fixed architecture, this one gives more results in all metrics, except in the number of entries in enclosed arms.
  • An analysis to the occurrences of each input sensor across generations shows that the most used ones are number 1 (right), 3 (left), and 2 (right, long-distance).
  • Used the genetic algorithm to measure how many sensory inputs are necessary to the network – three on average.

Thoughts

  • The authors do not provide strong hypotheses to why the preference of sensors on the right side. Although I can think why the preference for the sides rather than looking straight – possibly fear of danger coming from the peripherical view. It would be interesting to try to reproduce and hypothesize the reasons why this happens.
  • Over time, it becomes clear that the use of genetic algorithms can be a quite effective way of evolving and even helping to explain how a neural network works.

References

Floreano, D., Dürr, P., & Mattiussi, C. (2008). Neuroevolution: from architectures to learning. Evolutionary Intelligence, 1(1), 47–62. http://doi.org/10.1007/s12065-007-0002-4