Sebastian Risi

How novelty search escapes the deceptive trap of learning to learn

A major goal for researchers in neuroevolution is to evolve artificial neural networks (ANNs) that can learn during their lifetime. Such networks can adapt to changes in their environment that evolution on its own cannot anticipate. However, a profound problem with evolving adaptive systems is that if the impact of learning on the fitness of the agent is only marginal, then evolution is likely to produce individuals that do not exhibit the desired adaptive behavior. Instead, because it is easier at first to improve fitness without evolving the ability to learn, they are likely to exploit domain-dependent static (i.e. non-adaptive) heuristics. This paper proposes a way to escape the deceptive trap of static policies based on the novelty search algorithm, which opens up a new avenue in the evolution of adaptive systems because it can exploit the behavioral difference between learning and non-learning individuals. The main idea in novelty search is to abandon objective-based fitness and instead simply search only for novel behavior, which avoids deception entirely and has shown prior promising results in other domains. This paper shows that novelty search significantly outperforms fitness-based search in a tunably deceptive T-Maze navigation domain because it fosters the emergence of adaptive behavior.

Evolving plastic neural networks with novelty search

Biological brains can adapt and learn from past experience. Yet neuroevolution, that is, automatically creating artificial neural networks (ANNs) through evolutionary algorithms, has sometimes focused on static ANNs that cannot change their weights during their lifetime. A profound problem with evolving adaptive systems is that learning to learn is highly deceptive. Because it is easier at first to improve fitness without evolving the ability to learn, evolution is likely to exploit domain-dependent static (i.e., nonadaptive) heuristics. This article analyzes this inherent deceptiveness in a variety of different dynamic, reward-based learning tasks, and proposes a way to escape the deceptive trap of static policies based on the novelty search algorithm. The main idea in novelty search is to abandon objective-based fitness and instead simply search only for novel behavior, which avoids deception entirely. A series of experiments and an in-depth analysis show how behaviors that could potentially serve as a stepping stone to finding adaptive solutions are discovered by novelty search yet are missed by fitness-based search. The conclusion is that novelty search has the potential to foster the emergence of adaptive behavior in reward-based learning tasks, thereby opening a new direction for research in evolving plastic ANNs.

Indirectly encoding neural plasticity as a pattern of local rules

Biological brains can adapt and learn from past experience. In neuroevolution, i.e. evolving artificial neural networks (ANNs), one way that agents controlled by ANNs can evolve the ability to adapt is by encoding local learning rules. However, a significant problem with most such approaches is that local learning rules for every connection in the network must be discovered separately. This paper aims to show that learning rules can be effectively indirectly encoded by extending the Hypercube-based NeuroEvolution of Augmenting Topologies (HyperNEAT) method. Adaptive HyperNEAT is introduced to allow not only patterns of weights across the connectivity of an ANN to be generated by a function of its geometry, but also patterns of arbitrary learning rules. Several such adaptive models with different levels of generality are explored and compared. The long-term promise of the new approach is to evolve large-scale adaptive ANNs, which is a major goal for neuroevolution.

An enhanced hypercube-based encoding for evolving the placement, density, and connectivity of neurons

Intelligence in nature is the product of living brains, which are themselves the product of natural evolution. Although researchers in the field of neuroevolution (NE) attempt to recapitulate this process, artificial neural networks (ANNs) so far evolved through NE algorithms do not match the distinctive capabilities of biological brains. The recently introduced hypercube-based neuroevolution of augmenting topologies (HyperNEAT) approach narrowed this gap by demonstrating that the pattern of weights across the connectivity of an ANN can be generated as a function of its geometry, thereby allowing large ANNs to be evolved for high-dimensional problems. Yet the positions and number of the neurons connected through this approach must be decided a priori by the user and, unlike in living brains, cannot change during evolution. Evolvable-substrate HyperNEAT (ES-HyperNEAT), introduced in this article, addresses this limitation by automatically deducing the node geometry from implicit information in the pattern of weights encoded by HyperNEAT, thereby avoiding the need to evolve explicit placement. This approach not only can evolve the location of every neuron in the network, but also can represent regions of varying density, which means resolution can increase holistically over evolution. ES-HyperNEAT is demonstrated through multi-task, maze navigation, and modular retina domains, revealing that the ANNs generated by this new approach assume natural properties such as neural topography and geometric regularity. Also importantly, ES-HyperNEATs compact indirect encoding can be seeded to begin with a bias toward a desired class of ANN topographies, which facilitates the evolutionary search. The main conclusion is that ES-HyperNEAT significantly expands the scope of neural structures that evolution can discover.

A unified approach to evolving plasticity and neural geometry

An ambitious long-term goal for neuroevolution, which studies how artificial evolutionary processes can be driven to produce brain-like structures, is to evolve neurocontrollers with a high density of neurons and connections that can adapt and learn from past experience. Yet while neuroevolution has produced successful results in a variety of domains, the scale of natural brains remains far beyond reach. This paper unifies a set of advanced neuroevolution techniques into a new method called adaptive evolvable-substrate HyperNEAT, which is a step toward more biologically-plausible artificial neural networks (ANNs). The combined approach is able to fully determine the geometry, density, and plasticity of an evolving neuromodulated ANN. These complementary capabilities are demonstrated in a maze-learning task based on similar experiments with animals. The most interesting aspect of this investigation is that the emergent neural structures are beginning to acquire more natural properties, which means that neuroevolution can begin to pose new problems and answer deeper questions about how brains evolved that are ultimately relevant to the field of AI as a whole.

Single-unit pattern generators for quadruped locomotion

Legged robots can potentially venture beyond the limits of wheeled vehicles. While creating controllers for such robots by hand is possible, evolutionary algorithms are an alternative that can reduce the burden of hand-crafting robotic controllers. Although major evolutionary approaches to legged locomotion can generate oscillations through popular techniques such as continuous time recurrent neural networks (CTRNNs) or sinusoidal input, they typically face a challenge in maintaining long-term stability. The aim of this paper is to address this challenge by introducing an effective alternative based on a new type of neuron called a single-unit pattern generator (SUPG). The SUPG, which is indirectly encoded by a compositional pattern producing network (CPPN) evolved by HyperNEAT, produces a flexible temporal activation pattern that can be reset and repeated at any time through an explicit trigger input, thereby allowing it to dynamically recalibrate over time to maintain stability. The SUPG approach, which is compared to CTRNNs and sinusoidal input, is shown to produce natural-looking gaits that exhibit superior stability over time, thereby providing a new alternative for evolving oscillatory locomotion.

Evolving the placement and density of neurons in the hyperneat substrate

The Hypercube-based NeuroEvolution of Augmenting Topologies (HyperNEAT) approach demonstrated that the pattern of weights across the connectivity of an artificial neural network (ANN) can be generated as a function of its geometry, thereby allowing large ANNs to be evolved for high-dimensional problems. Yet it left to the user the question of where hidden nodes should be placed in a geometry that is potentially infinitely dense. To relieve the user from this decision, this paper introduces an extension called evolvable-substrate HyperNEAT (ES-HyperNEAT) that determines the placement and density of the hidden nodes based on a quadtree-like decomposition of the hypercube of weights and a novel insight about the relationship between connectivity and node placement. The idea is that the representation in HyperNEAT that encodes the pattern of connectivity across the ANN contains implicit information on where the nodes should be placed and can therefore be exploited to avoid the need to evolve explicit placement. In this paper, as a proof of concept, ES-HyperNEAT discovers working placements of hidden nodes for a simple navigation domain on its own, thereby eliminating the need to configure the HyperNEAT substrate by hand and suggesting the potential power of the new approach.

Encouraging reactivity to create robust machines

The robustness of animal behavior is unmatched by current machines, which often falter when exposed to unforeseen conditions. While animals are notably reactive to changes in their environment, machines often follow finely tuned yet inflexible plans. Thus, instead of the traditional approach of training such machines over many different unpredictable scenarios in detailed simulations (which is the most intuitive approach to inducing robustness), this work proposes to train machines to be reactive to their environment. The idea is that robustness may result not from detailed internal models or finely tuned control policies but from cautious exploratory behavior. Supporting this hypothesis, robots trained to navigate mazes with a reactive disposition prove more robust than those trained over many trials yet not rewarded for reactive behavior in both simulated tests and when embodied in real robots. The conclusion is that robustness may neither require an accurate model nor finely calibrated behavior.

Enhancing es-hyperneat to evolve more complex regular neural networks

The recently-introduced evolvable-substrate HyperNEAT algorithm (ES-HyperNEAT) demonstrated that the placement and density of hidden nodes in an artificial neural network can be determined based on implicit information in an infinite-resolution pattern of weights, thereby avoiding the need to evolve explicit placement. However, ES-HyperNEAT is computationally expensive because it must search the entire hypercube, and was shown only to match the performance of the original HyperNEAT in a simple benchmark problem. Iterated ES-HyperNEAT, introduced in this paper, helps to reduce computational costs by focusing the search on a sequence of two-dimensional cross-sections of the hypercube and therefore makes possible searching the hypercube at a finer resolution. A series of experiments and an analysis of the evolved networks show for the first time that iterated ES-HyperNEAT not only matches but outperforms original HyperNEAT in more complex domains because ES-HyperNEAT can evolve networks with limited connectivity, elaborate on existing network structure, and compensate for movement of information within the hypercube.

Confronting the challenge of learning a flexible neural controller for a diversity of morphologies

The ambulatory capabilities of legged robots offer the potential for access to dangerous and uneven terrain without a risk to human life. However, while machine learning has proven effective at training such robots to walk, a significant limitation of such approaches is that controllers trained for a specific robot are likely to fail when transferred to a robot with a slightly different morphology. This paper confronts this challenge with a novel strategy: Instead of training a controller for a particular quadruped morphology, it evolves a special function (through a method called HyperNEAT) that takes morphology as input and outputs an entire neural network controller fitted to the specific morphology. Once such a relationship is learned the output controllers are able to work on a diversity of different morphologies. Highlighting the unique potential of such an approach, in this paper a neural controller evolved for three different robot morphologies, which differ in the length of their legs, can interpolate to never-seen intermediate morphologies without any further training. Thus this work suggests a new research path towards learning controllers for whole ranges of morphologies: Instead of learning controllers themselves, it is possible to learn the relationship between morphology and control.