The use of autonomous robots for delivery of goods to customers is an exciting new way to provide a reliable and sustainable service. However, in the real world, autonomous robots still require human supervision for safety reasons. We tackle the realworld problem of optimizing autonomous robot timings to maximize deliveries, while ensuring that there are never too many robots running simultaneously so that they can be monitored safely. We assess the use of a recent hybrid machine-learningoptimization approach COIL (constrained optimization in learned latent space) and compare it with a baseline genetic algorithm for the purposes of exploring variations of this problem. We also investigate new methods for improving the speed and efficiency of COIL. We show that only COIL can find valid solutions where appropriate numbers of robots run simultaneously for all problem variations tested. We also show that when COIL has learned its latent representation, it can optimize 10% faster than the GA, making it a good choice for daily re-optimization of robots where delivery requests for each day are allocated to robots while maintaining safe numbers of robots running at once.

In this study, we investigate the effectiveness of using the PID (Proportional - Integral - Derivative) control loop factors for modifying response thresholds in a decentralized, non-communicating, threshold-based swarm. Each agent in our swarm has a set of four thresholds, each corresponding to a task the agent is capable of performing. The agent will act on a particular task if the stimulus is higher than its corresponding threshold. The ability to modify their thresholds allows the agents to specialize dynamically in response to task demands. Current approaches to dynamic thresholds typically use a learning and forgetting process to adjust thresholds. These methods are able to effectively specialize once, but can have difficulty re-specializing if the task demands change. Our approach, inspired by the PID control loop, alters the threshold values based on the current task demand value, the change in task demand, and the cumulative sum of previous task demands. We show that our PID-inspired method is scalable and outperforms fixed and current learning and forgetting response thresholds with non-changing, constant, and abrupt changes in task demand. This superior performance is due to the ability of our method to re-specialize repeatedly in response to changing task demands.

Physics simulations are a computational bottleneck in computer-aided design (CAD) optimization processes. Hence, in order to make accurate (computationally expensive) simulations feasible for use in design optimization, one requires either an optimization framework that is highly sample-efficient or fast data-driven proxies (surrogate models) for long running simulations. In this work, we leverage recent advances in optimization and artificial intelligence (AI) to address both of these potential solutions, in the context of designing an optimal unmanned underwater vehicle (UUV). We first investigate and compare the sample efficiency and convergence behavior of different optimization techniques with a standard computational fluid dynamics (CFD) solver in the optimization loop. We then develop a deep neural network (DNN) based surrogate model to approximate drag forces that would otherwise be computed via direct numerical simulation with the CFD solver. The surrogate model is in turn used in the optimization loop of the hull design. Our study finds that the Bayesian Optimization Lower Condition Bound (BO LCB) algorithm is the most sample-efficient optimization framework and has the best convergence behavior of those considered. Subsequently, we show that our DNN-based surrogate model predicts drag force on test data in tight agreement with CFD simulations, with a mean absolute percentage error (MAPE) of 1.85%. Combining these results, we demonstrate a two-orders-of-magnitude speedup (with comparable accuracy) for the design optimization process when the surrogate model is used. To our knowledge, this is the first study applying Bayesian optimization and DNN-based surrogate modeling to the problem of UUV design optimization, and we share our developments as open-source software.

Generalizability and stability are two key objectives for operating reinforcement learning (RL) agents in the real world. Designing RL algorithms that optimize these objectives can be a costly and painstaking process. This paper presents MetaPG, an evolutionary method for automated design of actor-critic loss functions. MetaPG explicitly optimizes for generalizability and performance, and implicitly optimizes the stability of both metrics. We initialize our loss function population with Soft Actor-Critic (SAC) and perform multi-objective optimization using fitness metrics encoding single-task performance, zero-shot generalizability to unseen environment configurations, and stability across independent runs with different random seeds. On a set of continuous control tasks from the Real-World RL Benchmark Suite, we find that our method, using a single environment during evolution, evolves algorithms that improve upon SAC's performance and generalizability by 4% and 20%, respectively, and reduce instability up to 67%. Then, we scale up to more complex environments from the Brax physics simulator and replicate generalizability tests encountered in practical settings, such as different friction coefficients. MetaPG evolves algorithms that can obtain 10% better generalizability without loss of performance within the same meta-training environment and obtain similar results to SAC when doing cross-domain evaluations in other Brax environments. The evolution results are interpretable; by analyzing the structure of the best algorithms we identify elements that help optimizing certain objectives, such as regularization terms for the critic loss.

Computational creativity has contributed heavily to abstract art in modern era, allowing artists to create high quality, abstract two dimension (2D) arts with a high level of controllability and expressibility. However, even with computational approaches that have promising result in making concrete 3D art, computationally addressing abstract 3D art with high-quality and controllability remains an open question. To fill this gap, we propose to explore computational creativity in making abstract 3D art by bridging evolution strategies (ES) and 3D rendering through customizable parameterization of scenes. We demonstrate that our approach is capable of placing semi-transparent triangles in 3D scenes that, when viewed from specified angles, render into films that look like artists' specification expressed in natural language. This provides a new way for the artist to easily express creativity ideas for abstract 3D art. The supplementary material, which contains code, animation for all figures, and more examples, is here: https://es3dart.github.io/

The configuration of radar networks is a complex problem that is often performed manually by experts with the help of a simulator. Different numbers and types of radars as well as different locations that the radars shall cover give rise to different instances of the radar configuration problem. The exact modeling of these instances is complex, as the quality of the configurations depends on a large number of parameters, on internal radar processing, and on the terrains on which the radars need to be placed. Classic optimization algorithms can therefore not be applied to this problem, and we rely on "trial-and-error" black-box approaches. In this paper, we study the performances of 13 black-box optimization algorithms on 153 radar network configuration problem instances. The algorithms perform considerably better than human experts. Their ranking, however, depends on the budget of configurations that can be evaluated and on the elevation profile of the location. We therefore also investigate automated algorithm selection approaches. Our results demonstrate that a pipeline that extracts instance features from the elevation of the terrain performs on par with the classical, far more expensive approach that extracts features from the objective function.

Improving the fairness of machine learning models is a nuanced task that requires decision makers to reason about multiple, conflicting criteria. The majority of fair machine learning methods transform the error-fairness trade-off into a single objective problem with a parameter controlling the relative importance of error versus fairness. We propose instead to directly optimize the error-fairness tradeoff by using multi-objective optimization. We present a flexible framework for defining the fair machine learning task as a weighted classification problem with multiple cost functions. This framework is agnostic to the underlying prediction model as well as the metrics. We use multiobjective optimization to define the sample weights used in model training for a given machine learner, and adapt the weights to optimize multiple metrics of fairness and accuracy across a set of tasks. To reduce the number of optimized parameters, and to constrain their complexity with respect to population subgroups, we propose a novel meta-model approach that learns to map protected attributes to sample weights, rather than optimizing those weights directly. On a set of real-world problems, this approach outperforms current state-of-the-art methods by finding solution sets with preferable error/fairness trade-offs.

Foundation models are large machine learning models that can tackle various downstream tasks once trained on diverse and large-scale data, leading research trends in natural language processing, computer vision, and reinforcement learning. However, no foundation model exists for optical multilayer thin film structure inverse design. Current inverse design algorithms either fail to explore the global design space or suffer from low computational efficiency. To bridge this gap, we propose the Opto Generative Pretrained Transformer (OptoGPT). OptoGPT is a decoder-only transformer that auto-regressively generates designs based on specific spectrum targets. Trained on a large dataset of 10 million designs, our model demonstrates remarkable capabilities: 1) autonomous global design exploration by determining the number of layers (up to 20) while selecting the material (up to 18 distinct types) and thickness at each layer, 2) efficient designs for structural color, absorbers, filters, distributed brag reflectors, and Fabry-Pérot resonators within 0.1 seconds (comparable to simulation speeds), 3) the ability to output diverse designs, and 4) seamless integration of user-defined constraints. By overcoming design barriers regarding optical targets, material selections, and design constraints, OptoGPT can serve as a foundation model for optical multilayer thin film structure inverse design.

In recent work, Lissovoi, Oliveto, and Warwicker (Artificial Intelligence (2023)) proved that the Move Acceptance Hyper-Heuristic (MAHH) leaves the local optimum of the multimodal cliff benchmark with remarkable efficiency. With its $O(n^3)$ runtime, for almost all cliff widths $d,$ the MAHH massively outperforms the $\Theta(n^d)$ runtime of simple elitist evolutionary algorithms (EAs). For the most prominent multimodal benchmark, the jump functions, the given runtime estimates of $O(n^{2m} m^{-\Theta(m)})$ and $\Omega(2^{\Omega(m)})$, for gap size $m \ge 2$, are far apart and the real performance of MAHH is still an open question. In this work, we resolve this question. We prove that for any choice of the MAHH selection parameter~$p$, the expected runtime of the MAHH on a jump function with gap size $m = o(n^{1/2})$ is at least $\Omega(n^{2m-1} / (2m-1)!)$. This renders the MAHH much slower than simple elitist evolutionary algorithms with their typical $O(n^m)$ runtime. We also show that the MAHH with the global bit-wise mutation operator instead of the local one-bit operator optimizes jump functions in time $O(\min\{m n^m,\frac{n^{2m-1}}{m!\Omega(m)^{m-2}}\})$, essentially the minimum of the optimization times of the $(1+1)$ EA and the MAHH. This suggests that combining several ways to cope with local optima can be a fruitful approach.

Modern machine learning (ML) models are capable of impressive performances. However, their prowess is not due only to the improvements in their architecture and training algorithms but also to a drastic increase in computational power used to train them. Such a drastic increase led to a growing interest in distributed ML, which in turn made worker failures and adversarial attacks an increasingly pressing concern. While distributed byzantine resilient algorithms have been proposed in a differentiable setting, none exist in a gradient-free setting. The goal of this work is to address this shortcoming. For that, we introduce a more general definition of byzantine-resilience in ML - the \textit{model-consensus}, that extends the definition of the classical distributed consensus. We then leverage this definition to show that a general class of gradient-free ML algorithms - ($1,\lambda$)-Evolutionary Search - can be combined with classical distributed consensus algorithms to generate gradient-free byzantine-resilient distributed learning algorithms. We provide proofs and pseudo-code for two specific cases - the Total Order Broadcast and proof-of-work leader election.