It's possible to distribute the Internet to users via drones. However it is then necessary to place the drones according to the positions of the users. Moreover, the 5th Generation (5G) New Radio (NR) technology is designed to accommodate a wide range of applications and industries. The NGNM 5G White Paper \cite{5gwhitepaper} groups these vertical use cases into three categories: - enhanced Mobile Broadband (eMBB) - massive Machine Type Communication (mMTC) - Ultra-Reliable Low-latency Communication (URLLC). Partitioning the physical network into multiple virtual networks appears to be the best way to provide a customised service for each application and limit operational costs. This design is well known as \textit{network slicing}. Each drone must thus slice its bandwidth between each of the 3 user classes. This whole problem (placement + bandwidth) can be defined as an optimization problem, but since it is very hard to solve efficiently, it is almost always addressed by AI in the litterature. In my internship, I wanted to prove that viewing the problem as an optimization problem can still be useful, by building an hybrid solution involving on one hand AI and on the other optimization. I use it to achieve better results than approaches that use only AI, although at the cost of slightly larger (but still reasonable) computation times.

They enable efficient management of resources and network traffic, a definite advantage in the age of increasingly complex networks requiring dynamic management. By centralizing control and enabling flexible management, SDN offers new opportunities for network optimization. Nevertheless, fully realizing the potential of SDN requires the development of advanced and adaptive control methods. This article focuses on analyzing current methods of flexible control for SDN networks and presenting a solution to improve the efficiency and adaptability of network management. The approach presented is based on the application of machine learning, specifically the Reinforcement Learning (RL) [2]. This technique allows networks to make autonomous decisions based on previous experiences and dynamically changing conditions, which is similar to the way humans learn. The goal of the proposed solution is to not only increase network performance, but to improve its flexibility and real-time adaptability. The use of reinforcement learning enables dynamic and flexible control of network traffic, resulting in more efficient and responsive resource management [3]. The article reviews existing solutions and describes in detail the original approach developed in-its own, pointing out its potential benefits and implementation possibilities.

In recent years, significant progress has been made in multi-objective reinforcement learning (RL) research, which aims to balance multiple objectives by incorporating preferences for each objective. In most existing studies, specific preferences must be provided during deployment to indicate the desired policies explicitly. However, designing these preferences depends heavily on human prior knowledge, which is typically obtained through extensive observation of high-performing demonstrations with expected behaviors. In this work, we propose a simple yet effective offline adaptation framework for multi-objective RL problems without assuming handcrafted target preferences, but only given several demonstrations to implicitly indicate the preferences of expected policies. Additionally, we demonstrate that our framework can naturally be extended to meet constraints on safety-critical objectives by utilizing safe demonstrations, even when the safety thresholds are unknown. Empirical results on offline multi-objective and safe tasks demonstrate the capability of our framework to infer policies that align with real preferences while meeting the constraints implied by the provided demonstrations.

Kolmogorov-Arnold Networks (KAN) is an emerging neural network architecture in machine learning. It has greatly interested the research community about whether KAN can be a promising alternative of the commonly used Multi-Layer Perceptions (MLP). Experiments in various fields demonstrated that KAN-based machine learning can achieve comparable if not better performance than MLP-based methods, but with much smaller parameter scales and are more explainable. In this paper, we explore the incorporation of KAN into the actor and critic networks for offline reinforcement learning (RL). We evaluated the performance, parameter scales, and training efficiency of various KAN and MLP based conservative Q-learning (CQL) on the the classical D4RL benchmark for offline RL. Our study demonstrates that KAN can achieve performance close to the commonly used MLP with significantly fewer parameters. This provides us an option to choose the base networks according to the requirements of the offline RL tasks.

Black swan events are statistically rare occurrences that carry extremely high risks. A typical view of defining black swan events is heavily assumed to originate from an unpredictable time-varying environments; however, the community lacks a comprehensive definition of black swan events. To this end, this paper challenges that the standard view is incomplete and claims that high-risk, statistically rare events can also occur in unchanging environments due to human misperception of their value and likelihood, which we call as spatial black swan event. We first carefully categorize black swan events, focusing on spatial black swan events, and mathematically formalize the definition of black swan events. We hope these definitions can pave the way for the development of algorithms to prevent such events by rationally correcting human perception.

The present study introduces an innovative approach to the synthesis of Electroencephalogram (EEG) signals by integrating diffusion models with reinforcement learning. This integration addresses key challenges associated with traditional EEG data acquisition, including participant burden, privacy concerns, and the financial costs of obtaining high-fidelity clinical data. Our methodology enhances the generation of EEG signals with detailed temporal and spectral features, enriching the authenticity and diversity of synthetic datasets. The uniqueness of our approach lies in its capacity to concurrently model time-domain characteristics, such as waveform morphology, and frequency-domain features, including rhythmic brainwave patterns, within a cohesive generative framework. This is executed through the reinforcement learning model's autonomous selection of parameter update strategies, which steers the diffusion process to accurately reflect the complex dynamics inherent in EEG signals. We validate the efficacy of our approach using both the BCI Competition IV 2a dataset and a proprietary dataset, each collected under stringent experimental conditions. Our results indicate that the method preserves participant privacy by generating synthetic data that lacks biometric identifiers and concurrently improves the efficiency of model training by minimizing reliance on large annotated datasets. This research offers dual contributions: firstly, it advances EEG research by providing a novel tool for data augmentation and the advancement of machine learning algorithms; secondly, it enhances brain-computer interface technologies by offering a robust solution for training models on diverse and representative EEG datasets. Collectively, this study establishes a foundation for future investigations in neurological care and the development of tailored treatment protocols in neurorehabilitation.

Reinforcement Learning has revolutionized decision-making processes in dynamic environments, yet it often struggles with autonomously detecting and achieving goals without clear feedback signals. For example, in a Source Term Estimation problem, the lack of precise environmental information makes it challenging to provide clear feedback signals and to define and evaluate how the source's location is determined. To address this challenge, the Autonomous Goal Detection and Cessation (AGDC) module was developed, enhancing various RL algorithms by incorporating a self-feedback mechanism for autonomous goal detection and cessation upon task completion. Our method effectively identifies and ceases undefined goals by approximating the agent's belief, significantly enhancing the capabilities of RL algorithms in environments with limited feedback. To validate effectiveness of our approach, we integrated AGDC with deep Q-Network, proximal policy optimization, and deep deterministic policy gradient algorithms, and evaluated its performance on the Source Term Estimation problem. The experimental results showed that AGDC-enhanced RL algorithms significantly outperformed traditional statistical methods such as infotaxis, entrotaxis, and dual control for exploitation and exploration, as well as a non-statistical random action selection method. These improvements were evident in terms of success rate, mean traveled distance, and search time, highlighting AGDC's effectiveness and efficiency in complex, real-world scenarios.

Recent advancements in reinforcement learning (RL) have been fueled by large-scale data and deep neural networks, particularly for high-dimensional and complex tasks. Online RL methods like Proximal Policy Optimization (PPO) are effective in dynamic scenarios but require substantial real-time data, posing challenges in resource-constrained or slow simulation environments. Offline RL addresses this by pre-learning policies from large datasets, though its success depends on the quality and diversity of the data. This work proposes a framework that enhances PPO algorithms by incorporating a diffusion model to generate high-quality virtual trajectories for offline datasets. This approach improves exploration and sample efficiency, leading to significant gains in cumulative rewards, convergence speed, and strategy stability in complex tasks. Our contributions are threefold: we explore the potential of diffusion models in RL, particularly for offline datasets, extend the application of online RL to offline environments, and experimentally validate the performance improvements of PPO with diffusion models. These findings provide new insights and methods for applying RL to high-dimensional, complex tasks. Finally, we open-source our code at https://github.com/TianciGao/DiffPPO

Supervised learning approaches to offline reinforcement learning, particularly those utilizing the Decision Transformer, have shown effectiveness in continuous environments and for sparse rewards. However, they often struggle with long-horizon tasks due to the high compounding error of auto-regressive models. To overcome this limitation, we go beyond next-token prediction and introduce Planning Tokens, which contain high-level, long time-scale information about the agent's future. Predicting dual time-scale tokens at regular intervals enables our model to use these long-horizon Planning Tokens as a form of implicit planning to guide its low-level policy and reduce compounding error. This architectural modification significantly enhances performance on long-horizon tasks, establishing a new state-of-the-art in complex D4RL environments. Additionally, we demonstrate that Planning Tokens improve the interpretability of the model's policy through the interpretable plan visualisations and attention map.

Single-cell organisms and various cell types use a range of motility modes when following a chemical gradient, but it is unclear which mode is best suited for different gradients. Here, we model directional decision-making in chemotactic amoeboid cells as a stimulus-dependent actin recruitment contest. Pseudopods extending from the cell body compete for a finite actin pool to push the cell in their direction until one pseudopod wins and determines the direction of movement. Our minimal model provides a quantitative understanding of the strategies cells use to reach the physical limit of accurate chemotaxis, aligning with data without explicit gradient sensing or cellular memory for persistence. To generalize our model, we employ reinforcement learning optimization to study the effect of pseudopod suppression, a simple but effective cellular algorithm by which cells can suppress possible directions of movement. Different pseudopod-based chemotaxis strategies emerge naturally depending on the environment and its dynamics. For instance, in static gradients, cells can react faster at the cost of pseudopod accuracy, which is particularly useful in noisy, shallow gradients where it paradoxically increases chemotactic accuracy. In contrast, in dynamics gradients, cells form \textit{de novo} pseudopods. Overall, our work demonstrates mechanical intelligence for high chemotaxis performance with minimal cellular regulation.