In this work, we attempt to reproduce the results of Enguehard (2023), which introduced ExtremalMask, a mask-based perturbation method for explaining time series data. We investigated the key claims of this paper, namely that (1) the model outperformed other models in several key metrics on both synthetic and real data, and (2) the model performed better when using the loss function of the preservation game relative to that of the deletion game. Although discrepancies exist, our results generally support the core of the original paper’s conclusions. Next, we interpret ExtremalMask’s outputs using new visualizations and metrics and discuss the insights each interpretation provides. Finally, we test whether ExtremalMask create out of distribution samples, and found the model does not exhibit this flaw on our tested synthetic dataset. Overall, our results support and add nuance to the original paper’s findings.

Due to a continuous increase in the amount of goods transported via water in the world, ports have become places of high maritime vessel traffic density. Because of this, the complexity of ports as systems has increased. Ports are pressured to constantly invest and to improve policies to increase the ports’ efficiency while maintaining a high level of safety and robustness. Proposed policies can be tested using simulation in a cost effective and safe way, but there is a lack of freely available, open-source harbour simulators for research and policy testing. In this paper, we present PySeidon, an open-source, modular, and generic port simulation framework written in Python. A proof-of-concept model of the Port of Rotterdam has been developed in PySeidon. We demonstrate the usefulness of this framework by performing scenario testing and comparing normal and anomalous agent behavior as well as two tugboat company policies with the model of the Port of Rotterdam.

Chain-of-Thought (CoT) generation has substantially improved the performance of Large Language Models (LLMs) on complex reasoning tasks, including code generation, data analysis, and exam-style question answering. Despite these advances, purely neural LLMs continue to struggle with elementary logical reasoning problems and lack the determinism, soundness, and reliability characteristic of symbolic reasoning systems. Conversely, classical symbolic methods such as SAT solving and Exact Cover guarantee correctness and completeness, but require problems to be expressed in highly specialized formal encodings, limiting their applicability to natural language inputs. In this work, we present a tightly integrated neuro-symbolic framework that bridges this gap by combining neural semantic parsing with deterministic constraint solving. Our approach leverages the relational extraction capabilities of modern LLMs to parse Zebra-style logic puzzles written in free-form text and translate the extracted constraints into structured tool calls. These function calls assemble a formally specified Exact Cover instance, which is subsequently solved by a symbolic solver to ensure logically consistent solutions. We conduct a comprehensive empirical evaluation across multiple parameter scales, post-training paradigms, and LLM families. The results on larger puzzles demonstrate that our hybrid approach consistently outperforms strong plain neural baselines, including CoT prompting, as well as recent neuro-symbolic methods.

Reinforcement learning (RL), one of the most successful methods for planning in stochastic environments, suffers from sample inefficiency, requiring extensive exploration of the environment to converge on good solutions. Additionally, most RL methods function as black boxes, limiting human intervention. This thesis attempts to tackle these problems and presents a method for learning temporal advice formulae to enhance the efficiency, quality, and safety of planning algorithms while maintaining transparency.

We use linear temporal logic on finite traces as a general framework for expressing advice. Inspired by previous works by Meli et al. and Ielo et al., we combine the existing research on learning time-independent advice for planners and inferring formulae from execution traces, to develop a unified method for learning temporal advice. We represent the temporal logic formulae as answer set programs and use the ILASP software for inductively learning them from execution traces. Unlike previous work, our approach tailors temporal logic formulae for guiding planning agents and accounts for partially observable and noisy domains. This integration enables automated advice generation, aiming to improve decision-making in automated planning.

We experimentally validate our approach in two environments: a simple fully observable gem pickup scenario and RockSample, which involves long planning horizons and partial observability. Our results demonstrate that generalizable temporal advice formulae can be learned from only a few examples, provided they are of high quality and clearly distinguish good from bad behavior.

Deep neural networks are powerful methods for modelling data and are increasingly deployed as part of safety-critical systems. Unfortunately, common neural networks evaluation methods draw conclusions which are often too optimistic for real-life scenarios. In this paper, we investigate verification, a method that ensures certain properties of neural networks. Specifically, natural language processing context is considered as research on verification of networks in this domain is scarce. We examine the performance of existing verification framework applications on networks performing sentiment classification as well as investigate piecewise linear activation function trade-offs. Moreover, latent space properties of many text representation techniques are investigated. Our empirical results show that the latent space induced by an autoencoder trained with a denoising adversarial objective is useful for verifying robustness of networks performing sentiment classification as well as interpeting results of verification tool outcomes.