Learning effective reinforcement learning (RL) policies to solve real-world complex tasks can be quite challenging without a high-fidelity simulation environment. In most cases, we are only given imperfect simulators with simplified dynamics, which inevitably lead to severe sim-to-real gaps in RL policy learning. The recently emerged field of offline RL provides another possibility to learn policies directly from pre-collected historical data. However, to achieve reasonable performance, existing offline RL algorithms need impractically large offline data with sufficient state-action space coverage for training. This brings up a new question: is it possible to combine learning from limited real data in offline RL and unrestricted exploration through imperfect simulators in online RL to address the drawbacks of both approaches? In this study, we propose the Dynamics-Aware Hybrid Offline-and-Online Reinforcement Learning (H2O) framework to provide an affirmative answer to this question. H2O introduces a dynamics-aware policy evaluation scheme, which adaptively penalizes the Q function learning on simulated state-action pairs with large dynamics gaps, while also simultaneously allowing learning from a fixed real-world dataset. Through extensive simulation and real-world tasks, as well as theoretical analysis, we demonstrate the superior performance of H2O against other cross-domain online and offline RL algorithms. H2O provides a brand new hybrid offline-and-online RL paradigm, which can potentially shed light on future RL algorithm design for solving practical real-world tasks.

Since personal computers became widely available in the consumer market, the amount of harmful content on the internet has significantly expanded. In simple terms, harmful content is anything online which causes a person distress or harm. It may include hate speech, violent content, threats, non-hope speech, etc. The online content must be positive, uplifting and supportive. Over the past few years, many studies have focused on solving this problem through hate speech detection, but very few focused on identifying hope speech. This paper discusses various machine learning approaches to identify a sentence as Hope Speech, Non-Hope Speech, or a Neutral sentence. The dataset used in the study contains English YouTube comments and is released as a part of the shared task "EACL-2021: Hope Speech Detection for Equality, Diversity, and Inclusion". Initially, the dataset obtained from the shared task had three classes: Hope Speech, non-Hope speech, and not in English; however, upon deeper inspection, we discovered that dataset relabeling is required. A group of undergraduates was hired to help perform the entire dataset's relabeling task. We experimented with conventional machine learning models (such as Na\"ive Bayes, logistic regression and support vector machine) and pre-trained models (such as BERT) on relabeled data. According to the experimental results, the relabeled data has achieved a better accuracy for Hope speech identification than the original data set.

Fairness in machine learning is crucial when individuals are subject to automated decisions made by models in high-stake domains. Organizations that employ these models may also need to satisfy regulations that promote responsible and ethical A.I. While fairness metrics relying on comparing model error rates across subpopulations have been widely investigated for the detection and mitigation of bias, fairness in terms of the equalized ability to achieve recourse for different protected attribute groups has been relatively unexplored. We present a novel formulation for training neural networks that considers the distance of data points to the decision boundary such that the new objective: (1) reduces the average distance to the decision boundary between two groups for individuals subject to a negative outcome in each group, i.e. the network is more fair with respect to the ability to obtain recourse, and (2) increases the average distance of data points to the boundary to promote adversarial robustness. We demonstrate that training with this loss yields more fair and robust neural networks with similar accuracies to models trained without it. Moreover, we qualitatively motivate and empirically show that reducing recourse disparity across groups also improves fairness measures that rely on error rates. To the best of our knowledge, this is the first time that recourse capabilities across groups are considered to train fairer neural networks, and a relation between error rates based fairness and recourse based fairness is investigated.

Explainable machine learning seeks to provide various stakeholders with insights into model behavior via feature importance scores, counterfactual explanations, and influential samples, among other techniques. Recent advances in this line of work, however, have gone without surveys of how organizations are using these techniques in practice. This study explores how organizations view and use explainability for stakeholder consumption. We find that the majority of deployments are not for end users affected by the model but for machine learning engineers, who use explainability to debug the model itself. There is a gap between explainability in practice and the goal of public transparency, since explanations primarily serve internal stakeholders rather than external ones. Our study synthesizes the limitations with current explainability techniques that hamper their use for end users. To facilitate end user interaction, we develop a framework for establishing clear goals for explainability, including a focus on normative desiderata.

As artificial intelligence plays an increasingly important role in our society, there are ethical and moral obligations for both businesses and researchers to ensure that their machine learning models are designed, deployed, and maintained responsibly. These models need to be rigorously audited for fairness, robustness, transparency, and interpretability. A variety of methods have been developed that focus on these issues in isolation, however, managing these methods in conjunction with model development can be cumbersome and timeconsuming. In this paper, we introduce a unified and model-agnostic approach to address these issues: Counterfactual Explanations for Robustness, Transparency, Interpretability, and Fairness of Artificial Intelligence models (CERTIFAI). Unlike previous methods in this domain, CERTIFAI is a general tool that can be applied to any black-box model and any type of input data. Given a model and an input instance, CERTIFAI uses a custom genetic algorithm to generate counterfactuals: instances close to the input that change the prediction of the model. We demonstrate how these counterfactuals can be used to examine issues of robustness, interpretability, transparency, and fairness. Additionally, we introduce CERScore, the first black-box model robustness score that performs comparably to methods that have access to model internals.