Giuseppe Cuccu

Evolving large-scale neural networks for vision-based reinforcement learning

The idea of using evolutionary computation to train artificial neural networks, or neuroevolution (NE), for reinforcement learning (RL) tasks has now been around for over 20 years. However, as RL tasks become more challenging, the networks required become larger, as do their genomes. But, scaling NE to large nets (i.e. tens of thousands of weights) is infeasible using direct encodings that map genes one-to-one to network components. In this paper, we scale-up our compressed network encoding where network weight matrices are represented indirectly as a set of Fourier-type coefficients, to tasks that require very-large networks due to the high-dimensionality of their input space. The approach is demonstrated successfully on two reinforcement learning tasks in which the control networks receive visual input: (1) a vision-based version of the octopus control task requiring networks with over 3 thousand weights, and (2) a version of the TORCS driving game where networks with over 1 million weights are evolved to drive a car around a track using video images from the drivers perspective.

Novelty-based restarts for evolution strategies

A major limitation in applying evolution strategies to black box optimization is the possibility of convergence into bad local optima. Many techniques address this problem, mostly through restarting the search. However, deciding the new start location is nontrivial since neither a good location nor a good scale for sampling a random restart position are known. A black box search algorithm can nonetheless obtain some information about this location and scale from past exploration. The method proposed here makes explicit use of such experience, through the construction of an archive of novel solutions during the run. Upon convergence, the most “novel” individual found so far is used to position the new start in the least explored region of the search space, actively looking for a new basin of attraction. We demonstrate the working principle of the method on two multi-modal test problems.

Intrinsically motivated neuroevolution for vision-based reinforcement learning

Neuroevolution, the artificial evolution of neural networks, has shown great promise on continuous reinforcement learning tasks that require memory. However, it is not yet directly applicable to realistic embedded agents using high-dimensional (e.g. raw video images) inputs, requiring very large networks. In this paper, neuroevolution is combined with an unsupervised sensory pre-processor or compressor that is trained on images generated from the environment by the population of evolving recurrent neural network controllers. The compressor not only reduces the input cardinality of the controllers, but also biases the search toward novel controllers by rewarding those controllers that discover images that it reconstructs poorly. The method is successfully demonstrated on a vision-based version of the well-known mountain car benchmark, where controllers receive only single high-dimensional visual images of the environment, from a third-person perspective, instead of the standard two-dimensional state vector which includes information about velocity.

Variable size population for dynamic optimization with genetic programming

A new model of Genetic Programming with variable size population is presented in this paper and applied to the reconstruction of target functions in dynamic environments (i.e. problems where target functions change with time). The suitability of this model is tested on a set of benchmarks based on some well known symbolic regression problems. Experimental results confirm that our variable size population model finds solutions of the same quality as the ones found by standard Genetic Programming, but with a smaller amount of computational effort.

Block diagonal natural evolution strategies

The Natural Evolution Strategies (NES) family of search algorithms have been shown to be efficient black-box optimizers, but the most powerful version xNES does not scale to problems with more than a few hundred dimensions. And the scalable variant, SNES, potentially ignores important correlations between parameters. This paper introduces Block Diagonal NES (BD-NES), a variant of NES which uses a block diagonal covariance matrix. The resulting update equations are computationally effective on problems with much higher dimensionality than their full-covariance counterparts, while retaining faster convergence speed than methods that ignore covariance information altogether. The algorithm has been tested on the Octopus-arm benchmark, and the experiments section presents performance statistics showing that BD-NES achieves better performance than SNES on networks that are too large to be optimized by xNES.

Reconstructing Dynamic Target Functions by Means of Genetic Programming Using Variable Population Size

Dynamic environments are becoming more and more popular in many applicative domains. A large amount of literature has appeared to date dealing with the problem of tracking the extrema of dynamically changing target functions, but relatively few material has been produced on the problem of reconstructing the shape, or more generally finding the equation, of dynamically changing target functions. Nevertheless, in many applicative domains, reaching this goal would have an extremely important impact. It is the case, for instance, of complex systems modelling, like for instance biological systems or systems of biochemical reactions, where one is generally interested in understanding what’s going on in the system over time, rather than following the extrema of some target functions. Last but not least, we also believe that being able to reach this goal would help researchers to have a useful insight on the reasons that cause the change in the system over time, or at least the pattern of this modification. This paper is intended as a first preliminary step in the attempt to fill this gap. We show that genetic programming with variable population size is able to adapt to the environment modifications much faster (i.e. using a noteworthy smaller amount of computational effort) than standard genetic programming using fixed population size. The suitability of this model is tested on a set of benchmarks based on some well known symbolic regression problems.