From 2017 to 2018 the number of scientific publications found via PubMed search using the keyword "Machine Learning" increased by 46% (4,317 to 6,307). The results of studies involving machine learning, artificial intelligence (AI), and big data have captured the attention of healthcare practitioners, healthcare managers, and the public at a time when Western medicine grapples with unmitigated cost increases and public demands for accountability. The complexity involved in healthcare applications of machine learning and the size of the associated data sets has afforded many researchers an uncontested opportunity to satisfy these demands with relatively little oversight. In a recent Nature Medicine article, "The Artificial Intelligence Clinician learns optimal treatment strategies for sepsis in intensive care," Komorowski and his coauthors propose methods to train an artificial intelligence clinician to treat sepsis patients with vasopressors and IV fluids. In this post, we will closely examine the claims laid out in this paper. In particular, we will study the individual treatment profiles suggested by their AI Clinician to gain insight into how their AI Clinician intends to treat patients on an individual level.

In many finite horizon episodic reinforcement learning (RL) settings, it is desirable to optimize for the undiscounted return - in settings like Atari, for instance, the goal is to collect the most points while staying alive in the long run. Yet, it may be difficult (or even intractable) mathematically to learn with this target. As such, temporal discounting is often applied to optimize over a shorter effective planning horizon. This comes at the cost of potentially biasing the optimization target away from the undiscounted goal. In settings where this bias is unacceptable - where the system must optimize for longer horizons at higher discounts - the target of the value function approximator may increase in variance leading to difficulties in learning. We present an extension of temporal difference (TD) learning, which we call TD($\Delta$), that breaks down a value function into a series of components based on the differences between value functions with smaller discount factors. The separation of a longer horizon value function into these components has useful properties in scalability and performance. We discuss these properties and show theoretic and empirical improvements over standard TD learning in certain settings.

Despite many algorithmic advances, our theoretical understanding of practical distributional reinforcement learning methods remains limited. One exception is Rowland et al. (2018)'s analysis of the C51 algorithm in terms of the Cram\'er distance, but their results only apply to the tabular setting and ignore C51's use of a softmax to produce normalized distributions. In this paper we adapt the Cram\'er distance to deal with arbitrary vectors. From it we derive a new distributional algorithm which is fully Cram\'er-based and can be combined to linear function approximation, with formal guarantees in the context of policy evaluation. In allowing the model's prediction to be any real vector, we lose the probabilistic interpretation behind the method, but otherwise maintain the appealing properties of distributional approaches. To the best of our knowledge, ours is the first proof of convergence of a distributional algorithm combined with function approximation. Perhaps surprisingly, our results provide evidence that Cram\'er-based distributional methods may perform worse than directly approximating the value function.

The state is the output of the vehicle's perception network (i.e. This has been suggested as a way to simulate road users' behaviour. What do you think of this approach?

The course rating is 4.5.

Actor-critic algorithms learn an explicit policy (actor), and an accompanying value function (critic). The actor performs actions in the environment, while the critic evaluates the actor's current policy. However, despite their stability and promising convergence properties, current actor-critic algorithms do not outperform critic-only ones in practice. We believe that the fact that the critic learns Q^pi, instead of the optimal Q-function Q*, prevents state-of-the-art robust and sample-efficient off-policy learning algorithms from being used. In this paper, we propose an elegant solution, the Actor-Advisor architecture, in which a Policy Gradient actor learns from unbiased Monte-Carlo returns, while being shaped (or advised) by the Softmax policy arising from an off-policy critic. The critic can be learned independently from the actor, using any state-of-the-art algorithm. Being advised by a high-quality critic, the actor quickly and robustly learns the task, while its use of the Monte-Carlo return helps overcome any bias the critic may have. In addition to a new Actor-Critic formulation, the Actor-Advisor, a method that allows an external advisory policy to shape a Policy Gradient actor, can be applied to many other domains. By varying the source of advice, we demonstrate the wide applicability of the Actor-Advisor to three other important subfields of RL: safe RL with backup policies, efficient leverage of domain knowledge, and transfer learning in RL. Our experimental results demonstrate the benefits of the Actor-Advisor compared to state-of-the-art actor-critic methods, illustrate its applicability to the three other application scenarios listed above, and show that many important challenges of RL can now be solved using a single elegant solution.

Reinforcement learning (RL) agents have traditionally been tasked with maximizing the value function of a Markov decision process (MDP), either in continuous settings, with fixed discount factor $\gamma < 1$, or in episodic settings, with $\gamma = 1$. While this has proven effective for specific tasks with well-defined objectives (e.g., games), it has never been established that fixed discounting is suitable for general purpose use (e.g., as a model of human preferences). This paper characterizes rationality in sequential decision making using a set of seven axioms and arrives at a form of discounting that generalizes traditional fixed discounting. In particular, our framework admits a state-action dependent "discount" factor that is not constrained to be less than 1, so long as there is eventual long run discounting. Although this broadens the range of possible preference structures in continuous settings, we show that there exists a unique "optimizing MDP" with fixed $\gamma < 1$ whose optimal value function matches the true utility of the optimal policy, and we quantify the difference between value and utility for suboptimal policies. Our work can be seen as providing a normative justification for (a slight generalization of) Martha White's RL task formalism (2017) and other recent departures from the traditional RL, and is relevant to task specification in RL, inverse RL and preference-based RL.

Multi-step methods such as Retrace($\lambda$) and $n$-step $Q$-learning have become a crucial component of modern deep reinforcement learning agents. These methods are often evaluated as a part of bigger architectures and their evaluations rarely include enough samples to draw statistically significant conclusions about their performance. This type of methodology makes it difficult to understand how particular algorithmic details of multi-step methods influence learning. In this paper we combine the $n$-step action-value algorithms Retrace, $Q$-learning, Tree Backup, Sarsa, and $Q(\sigma)$ with an architecture analogous to DQN. We test the performance of all these algorithms in the mountain car environment; this choice of environment allows for faster training times and larger sample sizes. We present statistical analyses on the effects of the off-policy correction, the backup length parameter $n$, and the update frequency of the target network on the performance of these algorithms. Our results show that (1) using off-policy correction can have an adverse effect on the performance of Sarsa and $Q(\sigma)$; (2) increasing the backup length $n$ consistently improved performance across all the different algorithms; and (3) the performance of Sarsa and $Q$-learning was more robust to the effect of the target network update frequency than the performance of Tree Backup, $Q(\sigma)$, and Retrace in this particular task.

This paper motivates and develops source traces for temporal difference (TD) learning in the tabular setting. Source traces are like eligibility traces, but model potential histories rather than immediate ones. This allows TD errors to be propagated to potential causal states and leads to faster generalization. Source traces can be thought of as the model-based, backward view of successor representations (SR), and share many of the same benefits. This view, however, suggests several new ideas. First, a TD($\lambda$)-like source learning algorithm is proposed and its convergence is proven. Then, a novel algorithm for learning the source map (or SR matrix) is developed and shown to outperform the previous algorithm. Finally, various approaches to using the source/SR model are explored, and it is shown that source traces can be effectively combined with other model-based methods like Dyna and experience replay.

Training intelligent agents through reinforcement learning is a notoriously unstable procedure. Massive parallelization on GPUs and distributed systems has been exploited to generate a large amount of training experiences and consequently reduce instabilities, but the success of training remains strongly influenced by the choice of the hyperparameters. To overcome this issue, we introduce HyperTrick, a new metaoptimization algorithm, and show its effective application to tune hyperparameters in the case of deep reinforcement learning, while learning to play different Atari games on a distributed system. Our analysis provides evidence of the interaction between the identification of the optimal hyperparameters and the learned policy, that is typical of the case of metaoptimization for deep reinforcement learning. When compared with state-of-the-art metaoptimization algorithms, HyperTrick is characterized by a simpler implementation and it allows learning similar policies, while making a more effective use of the computational resources in a distributed system.