With the advent of foundation models, prompt tuning has positioned itself as an important technique for directing model behaviors and eliciting desired responses. Prompt tuning regards selecting appropriate keywords included into the input, thereby adapting to the downstream task without adjusting or fine-tuning the model parameters. There is a wide range of work in prompt tuning, from approaches that directly harness the backpropagated gradient signals from the model, to those employing black-box optimization such as reinforcement learning (RL) methods. Our primary focus is on RLPrompt, which aims to find optimal prompt tokens leveraging soft Q-learning. While the results show promise, we have observed that the prompts frequently appear unnatural, which impedes their interpretability. We address this limitation by using sparse Tsallis entropy regularization, a principled approach to filtering out unlikely tokens from consideration. We extensively evaluate our approach across various tasks, including few-shot text classification, unsupervised text style transfer, and textual inversion from images. The results indicate a notable improvement over baselines, highlighting the efficacy of our approach in addressing the challenges of prompt tuning. Moreover, we show that the prompts discovered using our method are more natural and interpretable compared to those from other baselines.

One effective way to optimize the offloading process is by minimizing the transmission time. This is particularly true in a Vehicular Adhoc Network (VANET) where vehicles frequently download and upload High-definition (HD) map data which requires constant updates. This implies that latency and throughput requirements must be guaranteed by the wireless system. To achieve this, adjustable contention windows (CW) allocation strategies in the standard IEEE802.11p have been explored by numerous researchers. Nevertheless, their implementations demand alterations to the existing standard which is not always desirable. To address this issue, we proposed a Q-Learning algorithm that operates at the application layer. Moreover, it could be deployed in any wireless network thereby mitigating the compatibility issues. The solution has demonstrated a better network performance with relatively fewer optimization requirements as compared to the Deep Q Network (DQN) and Actor-Critic algorithms. The same is observed while evaluating the model in a multi-agent setup showing higher performance compared to the single-agent setup.

Social rewards shape human behavior. During development, a caregiver guides a learner's behavior towards culturally aligned goals and values. How do these behaviors persist and generalize when the caregiver is no longer present, and the learner must continue autonomously? Here, we propose a model of value internalization where social feedback trains an internal social reward (ISR) model that generates internal rewards when social rewards are unavailable. Through empirical simulations, we show that an ISR model prevents agents from unlearning socialized behaviors and enables generalization in out-of-distribution tasks. We characterize the implications of incomplete internalization, akin to "reward hacking" on the ISR. Additionally, we show that our model internalizes prosocial behavior in a multi-agent environment. Our work provides a foundation for understanding how humans acquire and generalize values and offers insights for aligning AI with human values.

Reinforcement learning from human feedback (RLHF) is a key driver of quality and safety in state-of-the-art large language models. Yet, a surprisingly simple and strong inference-time strategy is Best-of-N sampling that selects the best generation among N candidates. In this paper, we propose Best-of-N Distillation (BOND), a novel RLHF algorithm that seeks to emulate Best-of-N but without its significant computational overhead at inference time. Specifically, BOND is a distribution matching algorithm that forces the distribution of generations from the policy to get closer to the Best-of-N distribution. We use the Jeffreys divergence (a linear combination of forward and backward KL) to balance between mode-covering and mode-seeking behavior, and derive an iterative formulation that utilizes a moving anchor for efficiency. We demonstrate the effectiveness of our approach and several design choices through experiments on abstractive summarization and Gemma models. Aligning Gemma policies with BOND outperforms other RLHF algorithms by improving results on several benchmarks.

Financial portfolio management investment policies computed quantitatively by modern portfolio theory techniques like the Markowitz model rely on a set on assumptions that are not supported by data in high volatility markets. Hence, quantitative researchers are looking for alternative models to tackle this problem. Concretely, portfolio management is a problem that has been successfully addressed recently by Deep Reinforcement Learning (DRL) approaches. In particular, DRL algorithms train an agent by estimating the distribution of the expected reward of every action performed by an agent given any financial state in a simulator. However, these methods rely on Deep Neural Networks model to represent such a distribution, that although they are universal approximator models, they cannot explain its behaviour, given by a set of parameters that are not interpretable. Critically, financial investors policies require predictions to be interpretable, so DRL agents are not suited to follow a particular policy or explain their actions. In this work, we developed a novel Explainable Deep Reinforcement Learning (XDRL) approach for portfolio management, integrating the Proximal Policy Optimization (PPO) with the model agnostic explainable techniques of feature importance, SHAP and LIME to enhance transparency in prediction time. By executing our methodology, we can interpret in prediction time the actions of the agent to assess whether they follow the requisites of an investment policy or to assess the risk of following the agent suggestions. To the best of our knowledge, our proposed approach is the first explainable post hoc portfolio management financial policy of a DRL agent. We empirically illustrate our methodology by successfully identifying key features influencing investment decisions, which demonstrate the ability to explain the agent actions in prediction time.

We introduce a novel class of algorithms to efficiently approximate the unknown return distributions in policy evaluation problems from distributional reinforcement learning (DRL). The proposed distributional dynamic programming algorithms are suitable for underlying Markov decision processes (MDPs) having an arbitrary probabilistic reward mechanism, including continuous reward distributions with unbounded support being potentially heavy-tailed. For a plain instance of our proposed class of algorithms we prove error bounds, both within Wasserstein and Kolmogorov--Smirnov distances. Furthermore, for return distributions having probability density functions the algorithms yield approximations for these densities; error bounds are given within supremum norm. We introduce the concept of quantile-spline discretizations to come up with algorithms showing promising results in simulation experiments. While the performance of our algorithms can rigorously be analysed they can be seen as universal black box algorithms applicable to a large class of MDPs. We also derive new properties of probability metrics commonly used in DRL on which our quantitative analysis is based.

As the complexities of Dynamic Data Driven Applications Systems increase, preserving their resilience becomes more challenging. For instance, maintaining power grid resilience is becoming increasingly complicated due to the growing number of stochastic variables (such as renewable outputs) and extreme weather events that add uncertainty to the grid. Current optimization methods have struggled to accommodate this rise in complexity. This has fueled the growing interest in data-driven methods used to operate the grid, leading to more vulnerability to cyberattacks. One such disruption that is commonly discussed is the adversarial disruption, where the intruder attempts to add a small perturbation to input data in order to "manipulate" the system operation. During the last few years, work on adversarial training and disruptions on the power system has gained popularity. In this paper, we will first review these applications, specifically on the most common types of adversarial disruptions: evasion and poisoning disruptions. Through this review, we highlight the gap between poisoning and evasion research when applied to the power grid. This is due to the underlying assumption that model training is secure, leading to evasion disruptions being the primary type of studied disruption. Finally, we will examine the impacts of data poisoning interventions and showcase how they can endanger power grid resilience.

Despite tremendous progress, machine learning and deep learning still suffer from incomprehensible predictions. Incomprehensibility, however, is not an option for the use of (deep) reinforcement learning in the real world, as unpredictable actions can seriously harm the involved individuals. In this work, we propose a genetic programming framework to generate explanations for the decision-making process of already trained agents by imitating them with programs. Programs are interpretable and can be executed to generate explanations of why the agent chooses a particular action. Furthermore, we conduct an ablation study that investigates how extending the domain-specific language by using library learning alters the performance of the method. We compare our results with the previous state of the art for this problem and show that we are comparable in performance but require much less hardware resources and computation time.

When applying reinforcement learning from human feedback (RLHF), the reward is learned from data and, therefore, always has some error. It is common to mitigate this by regularizing the policy with KL divergence from a base model, with the hope that balancing reward with regularization will achieve desirable outcomes despite this reward misspecification. We show that when the reward function has light-tailed error, optimal policies under less restrictive KL penalties achieve arbitrarily high utility. However, if error is heavy-tailed, some policies obtain arbitrarily high reward despite achieving no more utility than the base model--a phenomenon we call catastrophic Goodhart. We adapt a discrete optimization method to measure the tails of reward models, finding that they are consistent with light-tailed error. However, the pervasiveness of heavy-tailed distributions in many real-world applications indicates that future sources of RL reward could have heavy-tailed error, increasing the likelihood of reward hacking even with KL regularization.

Software's pervasive impact and increasing reliance in the era of digital transformation raise concerns about vulnerabilities, emphasizing the need for software security. Fuzzy testing is a dynamic analysis software testing technique that consists of feeding faulty input data to a System Under Test (SUT) and observing its behavior. Specifically regarding black-box RESTful API testing, recent literature has attempted to automate this technique using heuristics to perform the input search and using the HTTP response status codes for classification. However, most approaches do not keep track of code coverage, which is important to validate the solution. This work introduces a black-box RESTful API fuzzy testing tool that employs Reinforcement Learning (RL) for vulnerability detection. The fuzzer operates via the OpenAPI Specification (OAS) file and a scenarios file, which includes information to communicate with the SUT and the sequences of functionalities to test, respectively. To evaluate its effectiveness, the tool was tested on the Petstore API. The tool found a total of six unique vulnerabilities and achieved 55\% code coverage.