We introduce RL-DARTS, one of the first applications of Differentiable Architecture Search (DARTS) in reinforcement learning (RL) to search for convolutional cells, applied to the Procgen benchmark. We outline the initial difficulties of applying neural architecture search techniques in RL, and demonstrate that by simply replacing the image encoder with a DARTS supernet, our search method is sample-efficient, requires minimal extra compute resources, and is also compatible with off-policy and on-policy RL algorithms, needing only minor changes in preexisting code. Surprisingly, we find that the supernet can be used as an actor for inference to generate replay data in standard RL training loops, and thus train end-to-end. Throughout this training process, we show that the supernet gradually learns better cells, leading to alternative architectures which can be highly competitive against manually designed policies, but also verify previous design choices for RL policies.

For continuing environments, reinforcement learning methods commonly maximize a discounted reward criterion with discount factor close to 1 in order to approximate the steady-state reward (the gain). However, such a criterion only considers the long-run performance, ignoring the transient behaviour. In this work, we develop a policy gradient method that optimizes the gain, then the bias (which indicates the transient performance and is important to capably select from policies with equal gain). We derive expressions that enable sampling for the gradient of the bias, and its preconditioning Fisher matrix. We further propose an algorithm that solves the corresponding bi-level optimization using a logarithmic barrier. Experimental results provide insights into the fundamental mechanisms of our proposal.

We study how different output layer types of a deep neural network learn and forget in continual learning settings. We describe the three factors affecting catastrophic forgetting in the output layer: (1) weights modifications, (2) interferences, and (3) projection drift. Our goal is to provide more insights into how different types of output layers can address (1) and (2). We also propose potential solutions and evaluate them on several benchmarks. We show that the best-performing output layer type depends on the data distribution drifts or the amount of data available. In particular, in some cases where a standard linear layer would fail, it is sufficient to change the parametrization and get significantly better performance while still training with SGD. Our results and analysis shed light on the dynamics of the output layer in continual learning scenarios and help select the best-suited output layer for a given scenario.

A long-standing challenge in artificial intelligence is lifelong learning. In lifelong learning, many tasks are presented in sequence and learners must efficiently transfer knowledge between tasks while avoiding catastrophic forgetting over long lifetimes. On these problems, policy reuse and other multi-policy reinforcement learning techniques can learn many tasks. However, they can generate many temporary or permanent policies, resulting in memory issues. Consequently, there is a need for lifetime-scalable methods that continually refine a policy library of a pre-defined size. This paper presents a first approach to lifetime-scalable policy reuse. To pre-select the number of policies, a notion of task capacity, the maximal number of tasks that a policy can accurately solve, is proposed. To evaluate lifetime policy reuse using this method, two state-of-the-art single-actor base-learners are compared: 1) a value-based reinforcement learner, Deep Q-Network (DQN) or Deep Recurrent Q-Network (DRQN); and 2) an actor-critic reinforcement learner, Proximal Policy Optimisation (PPO) with or without Long Short-Term Memory layer. By selecting the number of policies based on task capacity, D(R)QN achieves near-optimal performance with 6 policies in a 27-task MDP domain and 9 policies in an 18-task POMDP domain; with fewer policies, catastrophic forgetting and negative transfer are observed. Due to slow, monotonic improvement, PPO requires fewer policies, 1 policy for the 27-task domain and 4 policies for the 18-task domain, but it learns the tasks with lower accuracy than D(R)QN. These findings validate lifetime-scalable policy reuse and suggest using D(R)QN for larger and PPO for smaller library sizes. During their lifetime, animals may be subjected to a large number of unknown tasks. In some environmental conditions, nutritious food sources may be readily available, while in others they may be sparse, hidden, or even poisonous, and dangerous predators may roam in their vicinity. To address these challenging conditions, various behaviours must be selectively combined, such as avoidance, reward-seeking, or even fleeing. When direct perception provides limited or no cues about the current task, animals have to infer the task or use a strategy that works for many different tasks it may encounter. Therefore, a key challenge for learning challenging sequences of tasks is to find a limited number of strategies that work on the large domain of tasks that animals encounter over their lifetime. In artificial intelligence, variants of the above problem have been studied with investigations focusing on two aspects: transfer learning and catastrophic forgetting. Transfer learning is a process in which learners leverage the knowledge gained from a set of previously learned tasks with similar characteristics to a new task, whilst avoiding transferring knowledge that is not relevant [Taylor and Stone, 2009, Pan and Yang, 2010, Lazaric, 2013].

Unlike the conventional wisdom in statistical learning theory, the test error of a deep neural network (DNN) often demonstrates double descent: as the model complexity increases, it first follows a classical U-shaped curve and then shows a second descent. Through bias-variance decomposition, recent studies revealed that the bell-shaped variance is the major cause of model-wise double descent (when the DNN is widened gradually). This paper investigates epoch-wise double descent, i.e., the test error of a DNN also shows double descent as the number of training epoches increases. By extending the bias-variance analysis to epoch-wise double descent of the zero-one loss, we surprisingly find that the variance itself, without the bias, varies consistently with the test error. Inspired by this result, we propose a novel metric, optimization variance (OV), to measure the diversity of model updates caused by the stochastic gradients of random training batches drawn in the same iteration. OV can be estimated using samples from the training set only but correlates well with the (unknown) \emph{test} error, and hence early stopping may be achieved without using a validation set.

Domestic and service robots have the potential to transform industries such as health care and small-scale manufacturing, as well as the homes in which we live. However, due to the overwhelming variety of tasks these robots will be expected to complete, providing generic out-of-the-box solutions that meet the needs of every possible user is clearly intractable. To address this problem, robots must therefore not only be capable of learning how to complete novel tasks at run-time, but the solutions to these tasks must also be informed by the needs of the user. In this paper we demonstrate how behaviour trees, a well established control architecture in the fields of gaming and robotics, can be used in conjunction with natural language instruction to provide a robust and modular control architecture for instructing autonomous agents to learn and perform novel complex tasks. We also show how behaviour trees generated using our approach can be generalised to novel scenarios, and can be re-used in future learning episodes to create increasingly complex behaviours. We validate this work against an existing corpus of natural language instructions, demonstrate the application of our approach on both a simulated robot solving a toy problem, as well as two distinct real-world robot platforms which, respectively, complete a block sorting scenario, and a patrol scenario.

Peacefulness is a principal dimension of well-being for all humankind and is the way out of inequity and every single form of violence. Thus, its measurement has lately drawn the attention of researchers and policy-makers. During the last years, novel digital data streams have drastically changed the research in this field. In the current study, we exploit information extracted from Global Data on Events, Location, and Tone (GDELT) digital news database, to capture peacefulness through the Global Peace Index (GPI). Applying predictive machine learning models, we demonstrate that news media attention from GDELT can be used as a proxy for measuring GPI at a monthly level. Additionally, we use the SHAP methodology to obtain the most important variables that drive the predictions. This analysis highlights each country's profile and provides explanations for the predictions overall, and particularly for the errors and the events that drive these errors. We believe that digital data exploited by Social Good researchers, policy-makers, and peace-builders, with data science tools as powerful as machine learning, could contribute to maximize the societal benefits and minimize the risks to peacefulness.

The peaky behavior of CTC models is well known experimentally. However, an understanding about why peaky behavior occurs is missing, and whether this is a good property. We provide a formal analysis of the peaky behavior and gradient descent convergence properties of the CTC loss and related training criteria. Our analysis provides a deep understanding why peaky behavior occurs and when it is suboptimal. On a simple example which should be trivial to learn for any model, we prove that a feed-forward neural network trained with CTC from uniform initialization converges towards peaky behavior with a 100% error rate. Our analysis further explains why CTC only works well together with the blank label. We further demonstrate that peaky behavior does not occur on other related losses including a label prior model, and that this improves convergence.

Although ubiquitous in the sciences, histogram data have not received much attention by the Deep Learning community. Whilst regression and classification tasks for scalar and vector data are routinely solved by neural networks, a principled approach for estimating histogram labels as a function of an input vector or image is lacking in the literature. We present a dedicated method for Deep Learning-based histogram regression, which incorporates cross-bin information and yields distributions over possible histograms, expressed by $\tau$-quantiles of the cumulative histogram in each bin. The crux of our approach is a new loss function obtained by applying the pinball loss to the cumulative histogram, which for 1D histograms reduces to the Earth Mover's distance (EMD) in the special case of the median ($\tau = 0.5$), and generalizes it to arbitrary quantiles. We validate our method with an illustrative toy example, a football-related task, and an astrophysical computer vision problem. We show that with our loss function, the accuracy of the predicted median histograms is very similar to the standard EMD case (and higher than for per-bin loss functions such as cross-entropy), while the predictions become much more informative at almost no additional computational cost.

Multiple-Intent Inverse Reinforcement Learning (MI-IRL) seeks to find a reward function ensemble to rationalize demonstrations of different but unlabelled intents. Within the popular expectation maximization (EM) framework for learning probabilistic MI-IRL models, we present a warm-start strategy based on up-front clustering of the demonstrations in feature space. Our theoretical analysis shows that this warm-start solution produces a near-optimal reward ensemble, provided the behavior modes satisfy mild separation conditions. We also propose a MI-IRL performance metric that generalizes the popular Expected Value Difference measure to directly assesses learned rewards against the ground-truth reward ensemble. Our metric elegantly addresses the difficulty of pairing up learned and ground truth rewards via a min-cost flow formulation, and is efficiently computable. We also develop a MI-IRL benchmark problem that allows for more comprehensive algorithmic evaluations. On this problem, we find our MI-IRL warm-start strategy helps avoid poor quality local minima reward ensembles, resulting in a significant improvement in behavior clustering. Our extensive sensitivity analysis demonstrates that the quality of the learned reward ensembles is improved under various settings, including cases where our theoretical assumptions do not necessarily hold. Finally, we demonstrate the effectiveness of our methods by discovering distinct driving styles in a large real-world dataset of driver GPS trajectories.