This page gives an overview of EANT, a method to learn neural networks by evolutionary reinforcement learning

When a neural network is to be developed for a given problem, two aspects need to be considered:

1. What should the structure (or, topology) of the network be? More precisely, how many neural nodes does the network need in order to fulfil the demands of the given task, and what connections should be made between these nodes?
2. Given the structure of the neural network, what are the optimal values for its parameters? This includes the weights of the connections and possibly other parameters.

Traditionally the solution to aspect 1, the network's structure, is found by trial and error, or somehow determined beforehand using "intuition". Finding the solution to aspect 2, its parameters, is therefore the only aspect that is usually considered in the literature. It requires optimisation in a parameter space that can have a very high dimensionality—for difficult tasks it can be up to several hundred. This so-called "curse of dimensionality" is a significant obstacle in machine learning problems. When training a network's parameters by examples (e.g. supervised learning) it means that the number of training examples needed increases exponentially with the dimension of the parameter space. When using other methods of determining the parameters (e.g. reinforcement learning, as it is done here) the effects are different but equally detrimental. Most approaches for determining the parameters use backpropagation or similar methods that are, in effect, simple stochastic gradient descent optimisation algorithms.

In our opinion, the traditional approach described above has the following deficiencies:

1. The common approach to pre-design the network structure is difficult or even infeasible for complicated tasks. It can also result in overly complex networks if the designer cannot find a small structure that solves the task.
2. Determining the network parameters by local optimisation algorithms like gradient descent-type methods is impracticable for large problems. It is known from mathematical optimisation theory that these algorithms tend to get stuck in local minima. They only work well with very simple (e.g., convex) target functions or if an approximate solution is known beforehand.

In short, these methods lack generality and can therefore only be used to design neural networks for a small class of tasks. They are engineering-type approaches; there is nothing wrong with that if one needs to solve only a single, more or less constant problem (The No Free Lunch Theorem states that solutions that are specifically designed for a particular task always perform better at this task than more general methods. However, they perform worse on most or all other tasks, or if the task changes.) but it makes them unsatisfactory from a scientific point of view.

We believe that the best way to overcome these deficiencies is to replace these approaches by more general ones that are inspired by biology. Evolutionary theory tells us that the structure of the brain has been developed over a long period of time, starting from simple structures and getting more complex over time. The connections between biological neurons are modified by experience, i.e. learned and refined over a much shorter time span.

We have developed a method called EANT, Evolutionary Acquisition of Neural Topologies, that works in very much the same way to create a neural network as a solution to a given task. It is a very general learning algorithm that does not use any pre-defined knowledge of the task or the required solution. Instead, EANT uses evolutionary search methods on two levels:

1. In an outer optimisation loop called structural exploration new neural structures are developed by gradually adding new structure to an initially minimal network that is used as a starting point.
2. In an inner optimisation loop called structural exploitation the parameters of all currently considered structures are adjusted to maximise the performance of the networks on the given task.

EANT, originally developed by Yohannes Kassahun in the Cognitive Systems Group in Kiel during the years 2003-2006, has since been improved by replacing the structural exploitation loop by one that uses CMA-ES as its optimisation method, and by changing the way structural changes are made. Now it is in a state where it can, and has been, applied to very complex learning problems. It is being further developed under the lead of Nils T Siebel in the same research group.

The method has been tested with a simulation of a visual servoing setup. A robot arm with an attached hand is to be controlled by the neural network to move to a position where an object can be picked up. The only sensory data available to the network is visual data from a camera that overlooks the scene. EANT was used with a complete simulation of this visual servoing scenario to learn networks by evolutionary reinforcement learning.

The results obtained with EANT were very good, with an EANT controller of moderate size easily outperforming both traditional visual servoing and other neural network learning approaches, even though the traditional controller was given the true distance to the object in its Image Jacobian.

More details on EANT2, the improved version of EANT, and its application to robot visual servoing can be found in the relevant publications page here. More information on the original EANT can be found on Yohannes Kassahun's old publications page.

This page and all files in these subdirectories are Copyright © 2006-2008 Nils T Siebel, Christian-Albrechts-University of Kiel, Germany.