Research Interest

Prevailing research interests are situated within the fields of Evolutionary Robotics, Swarm-Robotics, Neuro-Evolution (combining the fields of Evolutionary Computation and Artificial Neural Networks) and statistical based machine learning techniques such as Deep Learning.

A current research goal is to formalize biologically inspired principles of selforganization, evolution and learning to control and direct the emergence of social phenomena such as cooperation and competition in simulated (agent-based) or physical (robotic) swarm systems (that is, systems comprising thousands of interacting components). Once design principles for encouraging and directing emergent phenomena in robot swarms (and more generally distributed systems) are formalized, then such emergent phenomena could be used as a problem solving tool to aid in solving a diverse range of complex tasks. For example, behavioral adaptation of agents in large scale multi-player computer games, nano-robots that adapt to accomplish various medical operations, automated design of energy efficient self-adapting power grids, and robotic swarm space exploration and cooperative construction of artificial habitats. This research falls under the purview of guided-self-organisation in simulated and physical swarm-based systems, which are types of adaptive artificial complex systems.


Current Research

Adapting Robotic Swarms for Collective Construction

This research is a preliminary step in producing adaptive (Neuro-Evolution) methods that autonomously adapt individual robot controllers (behaviours) and sensor configurations (morphologies) such that robot teams autonomously produce problem solving collective behaviours for high level user specified goals. For example, where desired structures (such as functional human habitats or complex machines) are specified by a user and robots adapt their individual behaviors in order to collectively build or self-assemble the desired structure.

This Evolutionary Robotics research falls within the larger taxonomy of cooperative multi-robot systems. The research objective is to investigate various neuro-evolution approaches for adapting robot controllers and morphologies (Watson and Nitschke, 2015) such that robot teams efficiently accomplish any permutation of a collective construction task in a given environment. Current research focuses on ascertaining the most appropriate method for the co-adaptation of behaviour and morphology for homogenous robot teams, given collective construction tasks of varying complexity.

This research is being done in collaboration with James Watson.

Evolution of Generalised Problem Solvers

A long time goal of artificial intelligence is to produce artificial brains capable of eliciting generalised problem solving behaviours equivalent to those observed in nature. Some research has focused on controller (behavioural) design methods which specifically aim to be general problem solvers across a broad range of task domains. However, an alternate approach is to demonstrate the efficacy of an existing controller design approach (such as neuro-evolution) as an effective problem solver for a range of task instances and then extract the methods underlying principles in order that a generalised problem solver, applicable to a broader range of tasks, can be derived. The objective of this research is to investigate the efficacy of comparative evolutionary neural network controller design methods to produce generalised problem solvers. Research to date has compared objective (fitness function) based, non-objective (behavioural diversity) based, and hybrid approaches (combining objective based search and behavioral diversity maintenance) to automatically derive neural controllers. Preliminary results have demonstrated the efficacy of the hybrid based approach for evolving generalised maze solving behaviours in an evolutionary robotics case study (Shorten and Nitschke, 2015). The current research focus is on investigating new methods to boost the evolvability (adaptability) of evolved controllers in order to facilitate the evolution of generalised problem solvers.

This research is being done in collaboration with David Shorten.

Complexification vs Simplification in Neuromorphic Robot Controllers

Neuroscience research has hypothesized that synaptic pruning as well as synaptic growth potentially plays a significant role in the plasticity and learning capacity of the animal brain. The research objective is to address the efficacy of this hypothesis using evolutionary robotics simulations to model adaptation via co-evolution (as occurs in nature). Predator-prey co-adaptation is modeled, that is co-evolution of robotic neuromorphic controllers, where predator robots must capture prey and prey robots must evade predators in dynamic environments. The first method tested is complexification, where agent behaviors are initially defined by simple Artificial Neural Network (ANN) controllers that become increasingly complex (via the addition of neurons and synaptic connections) over the course of an artificial evolution process. That us, ANN controllers evolve to be only as complex as they need to be in order to solve the given task. The second method is simplification, where robot behaviours are initially defined by large and complex ANN controllers that become smaller and simpler (due to synaptic pruning and the removal of neurons) throughout an artificial evolution process. Conversely, the key notion here is that ANN controllers evolve to be as simple as they need to be in order to solve the given task. The third method is a complexification-simplification hybrid that alternates between complexifying and simplifying ANN controllers based on varying task and environment requirements and constraints. Current research is testing these methods in a range of tasks and simulation environments including predator-prey evolutionary robotics simulations.

This research is being done in collaboration with Lucas Dreyer.

Automating the Design of Autonomous Collective Driving

Recently, there has been an increased interest in autonomous vehicles by various automotive companies. Future advanced transportation systems are envisaged to have thousands of autonomous vehicles which detect objects, avoid collisions, and predict accidents, whilst collectively traversing optimal paths through highways and road networks. Such a distributed control approach does not rely on any external control system and leaves behavioural autonomy to the individual vehicles.

The sensory configurations (number and type of sensors) and controllers (behaviours) of individual vehicles plays a critical role in the safe collective flow of autonomous traffic. However, current engineering design methods are not appropriate for the design of autonomous vehicles that must elicit a collective behaviour. That is, the safe and constant flow of traffic at given speeds for a vast range of roads and highways. In such cases, traffic systems can be viewed as complex systems where it is difficult, using traditional engineering methods, to ascertain what the exact sensory configuration and controller for each individual vehicle must be so as an optimal collective behaviour is synthesized. This research investigates the co-adaptation of the sensory configurations and controllers for autonomous cars using cooperative co-evolution approaches to automate the sensory design of vehicles and their behaviors in order that robust autonomous collective driving behaviors are synthesized.

This research is being done in collaboration with Allen Huang.

The Evolution of Music via Emotional Feedback

Algorithmic music composition is a centuries old idea, dating back to composition by Wolfgang Amadeus Mozart in 1792, where he developed a dice game by which musical motifs were randomly chosen from pre-designed lists to generate a complete piece. This research investigates the notion of automated music composition driven by artificial evolution, where the iterative evolutionary design process is guided by emotional (physiological) feedback of a human participant. There has been a significant amount of work on computer generated music, where there is either a human user guiding the iterative design process (aesthetic selection), or music generation is completely automated and guided by an objective function. Such objective functions attempt to encapsulate the subjective nature of a musical composition's broad appeal.

This research replaces such an objective function with physiological feedback that gauges one's emotional response to current music and thus guides algorithmic generation of future music. This creates an artificial evolution-physiological feedback loop that effectively guides the adaptation of a musical composition. The research objective is to ascertain if physiological metrics can be equated to emotion classifications that then guide an evolutionary algorithm's automatic generation of music specifically tailored to the moods and emotions of individual participants.

This research is being done in collaboration with Paul Cohen.

Collective Behaviour Transfer Learning in Multi-Robot Systems

The objective of this research is to derive methods that transfer knowledge between tasks in order to speed-up learning and to facilitate learning to solve a broad range of problems of increasing complexity. Using knowledge learned in a source task as the basis for learning in another more complex task is popularly known as transfer learning.

This research investigates transfer learning in multi-robot collective behaviour tasks. That is, learning to solve multi-robot tasks that require cooperative behaviour and then transferring the learned collective behaviour to a relatively difficult task that requires more complex form of collective behaviour. The collective behaviour learned in the source task is thus the starting point for learning a more complex collective behaviour in a target task. The problem of collective behaviour policy transfer is addressed via focusing on various representations of Neuro-Evolution (NE) methods. Representation, such as direct and indirect encoding of artificial neural network controllers is hypothesized as critical to successful collective behaviour transfer between tasks.

In the RoboCup soccer multi-robot task domain, comparative NE methods including NEAT (direct encoding) and HyperNEAT (indirect encoding) are being tested with the goal of ascertaining the most appropriate type of representation for adapting robot controllers and facilitating transfer learning. Experiments are testing collective behaviour policy transfer using the Keep-Away RoboCup soccer task with increasing levels of complexity between source and target tasks. Results indicate the efficacy of indirect encoding NE approaches for facilitating collective behavior transfer (Didi and Nitschke, 2016A), (Didi and Nitschke, 2016B). The current research focus is comparatively testing reinforcement learning transfer learning approaches with indirect NE approaches in a range of multi-robot collective behavior tasks.

This research is being done in collaboration with Sabre Didi.

Automating Application Protocol Identification in Networks

The proliferation of computer network users has, in recent years, placed strain on network resources, including bandwidth and number allocations. These issues are more apparent where connectivity is limited, such as in developing countries. The provisioning of services over these congested resources needs to be managed, ensuring a fair Quality of Experience (QoE) to consumers and producers alike. The Quality of Service (QoS) techniques used to manage such resources require constant revision and catering for new application protocols introduced to the network on a daily basis. The task of developing and refining these mechanisms is tedious, often requiring a significant amount of manual effort. These delays lead to a number of traffic flows being incorrectly identified (false positives) or remaining unclassified as they traverse a network. The inability to identify and manage these traffic flows may lead to excessive network congestion or render the network vulnerable to attack.

This research investigates methods for automating the development of real-time network traffic classifiers as well as increasing the accuracy of such automated methods compared to manual classification approaches. The Automated Pattern Identification and Classification (APIC) method is proposed to accomplish this objective. APIC comprises a combination of unsupervised and supervised machine learning algorithms to successfully detect previously unobserved application protocols, via training evolved artificial neural network classifiers to identify future instances of application protocols in real-time. Experimental results demonstrate APIC as classifying traffic flows more quickly and accurately than comparable methods (Goss and Nitschke, 2014).

This research is being done in collaboration with Ryan Goss.


Goss, R., and Nitschke, G. (2014). Automating Network Protocol Identification. In Biju, I and Nauman I., editors, Case Studies in Intelligent Computing - Achievements and Trends, pages 109-123. Taylor and Francis, New York, USA.

Didi, S. and Nitschke, G. (2016a). Hybridizing Novelty Search for Transfer Learning. In Proceedings of the IEEE Symposium Series on Computational Intelligence (IEEE SSCI 2016), to appear, Athens, Greece. IEEE Press.

Didi, S. and Nitschke, G. (2016b). Multi-Agent Behavior-Based Policy Transfer. In Proceedings of the European Conference on the Applications of Evolutionary Computation (Evostar 2016), pages 181-197, Porto, Portugal. Springer.

Shorten, D., and Nitschke, G. (2015). Evolving Generalised Maze Solvers. In Proceedings of the 18th European Conference on the Applications of Evolutionary Computation (Evostar 2015), pages 783-794, Copenhagen, Denmark, Springer.

Watson, J., and Nitschke, G. (2015). Deriving Minimal Sensory Configurations for Evolved Cooperative Robot Teams. In Proceedings of the IEEE Congress on Evolutionary Computation (IEEE CEC 2015), pages 3065-3071, IEEE Press, Sendai, Japan.