Artificial neural networks (ANNs) can be employed as controllers for robotic agents. Their structure is often complex, with many neurons and connections, especially when the robots have many sensors and actuators distributed across their bodies and/or when high expressive power is desirable. Pruning (removing neurons or connections) reduces the complexity of the ANN, thus increasing its energy efficiency, and has been reported to improve the generalization capability, in some cases. In addition, it is well-known that pruning in biological neural networks plays a fundamental role in the development of brains and their ability to learn. In this study, we consider the evolutionary optimization of neural controllers for the case study of Voxel-based soft robots, a kind of modular, bio-inspired soft robots, applying pruning during fitness evaluation. For a locomotion task, and for centralized as well as distributed controllers, we experimentally characterize the effect of different forms of pruning on after-pruning effectiveness, life-long effectiveness, adaptability to new terrains, and behavior. We find that incorporating some forms of pruning in neuroevolution leads to almost equally effective controllers as those evolved without pruning, with the benefit of higher robustness to pruning. We also observe occasional improvements in generalization ability.

We investigate artificial intelligence and machine learning methods for optimizing the adversarial behavior of agents in cybersecurity simulations. Our cybersecurity simulations integrate the modeling of agents launching Advanced Persistent Threats (APTs) with the modeling of agents using detection and mitigation mechanisms against APTs. This simulates the phenomenon of how attacks and defenses coevolve. The simulations and machine learning are used to search for optimal agent behaviors. The central question is: under what circumstances, is one training method more advantageous than another? We adapt and compare a variety of deep reinforcement learning (DRL), evolutionary strategies (ES) and Monte Carlo Tree Search methods within Connect 4, a baseline game environment, and on both a simulation supporting a simple APT threat model, SNAPT, as well as CyberBattleSim, an open-source cybersecurity simulation. Our results show that when attackers are trained by DRL and ES algorithms, as well as when they are trained with both algorithms being used in alternation, they are able to effectively choose complex exploits that thwart a defense. The algorithm that combines DRL and ES achieves the best comparative performance when attackers and defenders are simultaneously trained, rather than when each is trained against its non-learning counterpart.

Algorithmic bias arises in machine learning when models that may have reasonable overall accuracy are biased in favor of ‘good’ outcomes for one side of a sensitive category, for example gender or race. The bias will manifest as an underestimation of good outcomes for the under-represented minority. In a sense, we should not be surprised that a model might be biased when it has not been ‘asked’ not to be; reasonable accuracy can be achieved by ignoring the under-represented minority. A common strategy to address this issue is to include fairness as a component in the learning objective. In this paper, we consider including fairness as an additional criterion in model training and propose a multi-objective optimization strategy using Pareto Simulated Annealing that optimizes for both accuracy and underestimation bias. Our experiments show that this strategy can identify families of models with members representing different accuracy/fairness tradeoffs. We demonstrate the effectiveness of this strategy on two synthetic and two real-world datasets.

Reinforcement Learning (RL) methods often rely on gradient estimates to learn an optimal policy for control problems. These expensive computations result in long training times, a poor rate of convergence, and sample inefficiency when applied to real-world problems with a large state and action space. Evolutionary Computation (EC)-based techniques offer a gradient-free apparatus to train a deep neural network for RL problems. In this work, we leverage the benefits of EC and propose a novel variant of genetic algorithm called SP-GA which utilizes a species-inspired weight initialization strategy and trains a population of deep neural networks, each estimating the Q-function for the RL problem. Efficient encoding of a neural network that utilizes less memory is also proposed which provides an intuitive mechanism to apply Gaussian mutations and single-point crossover. The results on Atari 2600 games outline comparable performance with gradient-based algorithms like Deep Q-Network (DQN), Asynchronous Advantage Actor Critic (A3C), and gradient-free algorithms like Evolution Strategy (ES) and simple Genetic Algorithm (GA) while requiring far fewer hyperparameters to train. The algorithm also improved certain Key Performance Indicators (KPIs) when applied to a Remote Electrical Tilt (RET) optimization task in the telecommunication domain.