We consider online learning in episodic loopfree We propose a novel algorithm for the adversarial MDP Markov decision processes (MDPs), where model where the transition function is unknown to the the loss function can change arbitrarily between learner and the losses change arbitrarily over time. Our episodes, and the transition function is not known algorithm, UC-O-REPS, uses two important ingredients, to the learner. We show Õ(L X A T) regret the first is Online Mirror Descent (OMD) (Shalev-Shwartz, bound, where T is the number of episodes, X 2012) and the second is UCRL-2 (Auer et al., 2008). A is the state space, A is the action space, and L major challenge in this work is to handle convex performance is the length of each episode. Our online algorithm criteria, which model different ways of aggregating is implemented using entropic regularization the losses of each episode. In order to handle convex performance methodology, which allows to extend the criteria, we use the methodology of OMD, which is original adversarial MDP model to handle convex widely used for online convex optimization, and we implement performance criteria (different ways to aggregate it in the adversarial MDP setting. In order to overcome the losses of a single episode), as well as the unknown dynamics (stochastic transition function) improve previous regret bounds.

Many continuous control tasks have easily formulated objectives, yet using them directly as a reward in reinforcement learning (RL) leads to suboptimal policies. Therefore, many classical control tasks guide RL training using complex rewards, which require tedious hand-tuning. We automate the reward search with AutoRL, an evolutionary layer over standard RL that treats reward tuning as hyperparameter optimization and trains a population of RL agents to find a reward that maximizes the task objective. AutoRL, evaluated on four Mujoco continuous control tasks over two RL algorithms, shows improvements over baselines, with the the biggest uplift for more complex tasks. The video can be found at: \url{https://youtu.be/svdaOFfQyC8}.

Multi-output supervised learning aims to model input-output relationships from observations of inputoutput pairs whenever the output space is a vector of random variables. Multi-output classification and regression tasks have numerous applications in domains ranging from biology to multimedia, and recent applications in this area correspond to very high dimensional output spaces (Agrawal et al, 2013; Dekel and Shamir, 2010). Classification and regression trees (Breiman et al, 1984) are popular supervised learning methods that provide state-of-the-art performance when exploited in the context of ensemble methods, namely Random forests (Breiman, 2001; Geurts et al, 2006) and Boosting (Freund and Schapire, 1997; Friedman, 2001). Classification and regression trees can obviously be exploited to handle multi-output problems. The most straightforward way to address multi-output tasks is to apply standard single output methods separately and independently on each output. Although simple, this method, called binary relevance (Tsoumakas et al, 2009) in multi-label classification or single target (Spyromitros-Xioufis et al, 2012) in multi-output regression is often suboptimal as it does not exploit potential correlations that might exist between the outputs. Tree ensemble methods have however been explicitely extended by several authors to the joint prediction of multiple outputs (e.g., Segal, 1992; Blockeel et al, 2000). These extensions build a single tree to predict all outputs at once. They adapt the score measure used to assess splits during the tree growth to take into account all outputs and label each tree leaf with a vector of values, one for each output.

Due to the ease of modern data collection, applied statisticians often have access to a large set of covariates that they wish to relate to some observed outcome. Generalized linear models (GLMs) offer a particularly interpretable framework for such an analysis. In these high-dimensional problems, the number of covariates is often large relative to the number of observations, so we face non-trivial inferential uncertainty; a Bayesian approach allows coherent quantification of this uncertainty. Unfortunately, existing methods for Bayesian inference in GLMs require running times roughly cubic in parameter dimension, and so are limited to settings with at most tens of thousand parameters. We propose to reduce time and memory costs with a low-rank approximation of the data in an approach we call LR-GLM. When used with the Laplace approximation or Markov chain Monte Carlo, LR-GLM provides a full Bayesian posterior approximation and admits running times reduced by a full factor of the parameter dimension. We rigorously establish the quality of our approximation and show how the choice of rank allows a tunable computational-statistical trade-off. Experiments support our theory and demonstrate the efficacy of LR-GLM on real large-scale datasets.

Record companies invest billions of dollars in new talent around the globe each year. Gaining insight into what actually makes a hit song would provide tremendous benefits for the music industry. In this research we tackle this question by focussing on the dance hit song classification problem. A database of dance hit songs from 1985 until 2013 is built, including basic musical features, as well as more advanced features that capture a temporal aspect. A number of different classifiers are used to build and test dance hit prediction models. The resulting best model has a good performance when predicting whether a song is a "top 10" dance hit versus a lower listed position.

Health-related data is noisy and stochastic in implying the true physiological states of patients, limiting information contained in single-moment observations for sequential clinical decision making. We model patient-clinician interactions as partially observable Markov decision processes (POMDPs) and optimize sequential treatment based on belief states inferred from history sequence. To facilitate inference, we build a variational generative model and boost state representation with a recurrent neural network (RNN), incorporating an auxiliary loss from sequence auto-encoding. Meanwhile, we optimize a continuous policy of drug levels with an actor-critic method where policy gradients are obtained from a stablized off-policy estimate of advantage function, with the value of belief state backed up by parallel best-first suffix trees. We exploit our methodology in optimizing dosages of vasopressor and intravenous fluid for sepsis patients using a retrospective intensive care dataset and evaluate the learned policy with off-policy policy evaluation (OPPE). The results demonstrate that modelling as POMDPs yields better performance than MDPs, and that incorporating heuristic search improves sample efficiency.

Machine Learning techniques have become pervasive across a range of different applications, and are now widely used in areas as disparate as recidivism prediction, consumer credit-risk analysis and insurance pricing. The prevalence of machine learning techniques has raised concerns about the potential for learned algorithms to become biased against certain groups. Many definitions have been proposed in the literature, but the fundamental task of reasoning about probabilistic events is a challenging one, owing to the intractability of inference. The focus of this paper is taking steps towards the application of tractable models to fairness. Tractable probabilistic models have emerged that guarantee that conditional marginal can be computed in time linear in the size of the model. In particular, we show that sum product networks (SPNs) enable an effective technique for determining the statistical relationships between protected attributes and other training variables. If a subset of these training variables are found by the SPN to be independent of the training attribute then they can be considered `safe' variables, from which we can train a classification model without concern that the resulting classifier will result in disparate outcomes for different demographic groups. Our initial experiments on the `German Credit' data set indicate that this processing technique significantly reduces disparate treatment of male and female credit applicants, with a small reduction in classification accuracy compared to state of the art. We will also motivate the concept of "fairness through percentile equivalence", a new definition predicated on the notion that individuals at the same percentile of their respective distributions should be treated equivalently, and this prevents unfair penalisation of those individuals who lie at the extremities of their respective distributions.

Deep Gaussian processes (DGP) have appealing Bayesian properties, can handle variable-sized data, and learn deep features. Their limitation is that they do not scale well with the size of the data. Existing approaches address this using a deep random feature (DRF) expansion model, which makes inference tractable by approximating DGPs. However, DRF is not suitable for variable-sized input data such as trees, graphs, and sequences. We introduce the GP-DRF, a novel Bayesian model with an input layer of GPs, followed by DRF layers. The key advantage is that the combination of GP and DRF leads to a tractable model that can both handle a variable-sized input as well as learn deep long-range dependency structures of the data. We provide a novel efficient method to simultaneously infer the posterior of GP's latent vectors and infer the posterior of DRF's internal weights and random frequencies. Our experiments show that GP-DRF outperforms the standard GP model and DRF model across many datasets. Furthermore, they demonstrate that GP-DRF enables improved uncertainty quantification compared to GP and DRF alone, with respect to a Bhattacharyya distance assessment. Source code is available at https://github.com/IssamLaradji/GP_DRF.

We consider the problem of recovering a common latent source with independent components from multiple views. This applies to settings in which a variable is measured with multiple experimental modalities, and where the goal is to synthesize the disparate measurements into a single unified representation. We consider the case that the observed views are a nonlinear mixing of component-wise corruptions of the sources. When the views are considered separately, this reduces to nonlinear Independent Component Analysis (ICA) for which it is provably impossible to undo the mixing. We present novel identifiability proofs that this is possible when the multiple views are considered jointly, showing that the mixing can theoretically be undone using function approximators such as deep neural networks. In contrast to known identifiability results for nonlinear ICA, we prove that independent latent sources with arbitrary mixing can be recovered as long as multiple, sufficiently different noisy views are available.

We consider an agent who is involved in a Markov decision process and receives a vector of outcomes every round. Her objective is to maximize a global concave reward function on the average vectorial outcome. The problem models applications such as multi-objective optimization, maximum entropy exploration, and constrained optimization in Markovian environments. In our general setting where a stationary policy could have multiple recurrent classes, the agent faces a subtle yet consequential trade-off in alternating among different actions for balancing the vectorial outcomes. In particular, stationary policies are in general sub-optimal. We propose a no-regret algorithm based on online convex optimization (OCO) tools (Agrawal and Devanur 2014) and UCRL2 (Jaksch et al. 2010). Importantly, we introduce a novel gradient threshold procedure, which carefully controls the switches among actions to handle the subtle trade-off. By delaying the gradient updates, our procedure produces a non-stationary policy that diversifies the outcomes for optimizing the objective. The procedure is compatible with a variety of OCO tools.