On-policy reinforcement learning (RL) algorithms have high sample complexity while off-policy algorithms are difficult to tune. Merging the two holds the promise to develop efficient algorithms that generalize across diverse environments. It is however challenging in practice to find suitable hyper-parameters that govern this trade off. This paper develops a simple algorithm named P3O that interleaves off-policy updates with on-policy updates. P3O uses the effective sample size between the behavior policy and the target policy to control how far they can be from each other and does not introduce any additional hyper-parameters. Extensive experiments on the Atari-2600 and MuJoCo benchmark suites show that this simple technique is highly effective in reducing the sample complexity of state-of-the-art algorithms.

Hierarchy in reinforcement learning agents allows for control at multiple time scales yielding improved sample efficiency, the ability to deal with long time horizons and transferability of sub-policies to tasks outside the training distribution. It is often implemented as a master policy providing goals to a sub-policy. Ideally, we would like the goal-spaces to be learned, however, properties of optimal goal spaces still remain unknown and consequently there is no method yet to learn optimal goal spaces. Motivated by this, we systematically analyze how various modifications to the ground-truth goal-space affect learning in hierarchical models with the aim of identifying important properties of optimal goal spaces. Our results show that, while rotation of ground-truth goal spaces and noise had no effect, having additional unnecessary factors significantly impaired learning in hierarchical models.

We study online reinforcement learning for finite-horizon deterministic control systems with {\it arbitrary} state and action spaces. Suppose that the transition dynamics and reward function is unknown, but the state and action space is endowed with a metric that characterizes the proximity between different states and actions. We provide a surprisingly simple upper-confidence reinforcement learning algorithm that uses a function approximation oracle to estimate optimistic Q functions from experiences. We show that the regret of the algorithm after $K$ episodes is $O(HL(KH)^{\frac{d-1}{d}}) $ where $L$ is a smoothness parameter, and $d$ is the doubling dimension of the state-action space with respect to the given metric. We also establish a near-matching regret lower bound. The proposed method can be adapted to work for more structured transition systems, including the finite-state case and the case where value functions are linear combinations of features, where the method also achieve the optimal regret.

Financial portfolio management is one of the problems that are most frequently encountered in the investment industry. Nevertheless, it is not widely recognized that both Kelly Criterion and Risk Parity collapse into Mean Variance under some conditions, which implies that a universal solution to the portfolio optimization problem could potentially exist. In fact, the process of sequential computation of optimal component weights that maximize the portfolio's expected return subject to a certain risk budget can be reformulated as a discrete-time Markov Decision Process (MDP) and hence as a stochastic optimal control, where the system being controlled is a portfolio consisting of multiple investment components, and the control is its component weights. Consequently, the problem could be solved using model-free Reinforcement Learning (RL) without knowing specific component dynamics. By examining existing methods of both value-based and policy-based model-free RL for the portfolio optimization problem, we identify some of the key unresolved questions and difficulties facing today's portfolio managers of applying model-free RL to their investment portfolios.

Many real world tasks exhibit rich structure that is repeated across different parts of the state space or in time. In this work we study the possibility of leveraging such repeated structure to speed up and regularize learning. We start from the KL regularized expected reward objective which introduces an additional component, a default policy. Instead of relying on a fixed default policy, we learn it from data. But crucially, we restrict the amount of information the default policy receives, forcing it to learn reusable behaviours that help the policy learn faster. We formalize this strategy and discuss connections to information bottleneck approaches and to the variational EM algorithm. We present empirical results in both discrete and continuous action domains and demonstrate that, for certain tasks, learning a default policy alongside the policy can significantly speed up and improve learning.

We revisit residual algorithms in both model-free and model-based reinforcement learning settings. We propose the bidirectional target network technique to stabilize residual algorithms, yielding a residual version of DDPG that significantly outperforms vanilla DDPG in the DeepMind Control Suite benchmark. Moreover, we find the residual algorithm an effective approach to the distribution mismatch problem in model-based planning. Compared with the existing TD($k$) method, our residual-based method makes weaker assumptions about the model and yields a greater performance boost.

Is Artificial Intelligence taking over our jobs or can AI assist us to be better at our jobs? Whichever way you look at it, you can't deny how far we've come. These days we hear things like Reinforcement Learning and Machine Learning. We speak to Dr Marko Kesti, CEO of Playgain and Research Director at the University of Lapland in Finland to talk about the quality of life in the present and the future with artificial intelligence.

Air traffic control is a real-time safety-critical decision making process in highly dynamic and stochastic environments. In today's aviation practice, a human air traffic controller monitors and directs many aircraft flying through its designated airspace sector. With the fast growing air traffic complexity in traditional (commercial airliners) and low-altitude (drones and eVTOL aircraft) airspace, an autonomous air traffic control system is needed to accommodate high density air traffic and ensure safe separation between aircraft. We propose a deep multi-agent reinforcement learning framework that is able to identify and resolve conflicts between aircraft in a high-density, stochastic, and dynamic en-route sector with multiple intersections and merging points. The proposed framework utilizes an actor-critic model, A2C that incorporates the loss function from Proximal Policy Optimization (PPO) to help stabilize the learning process. In addition we use a centralized learning, decentralized execution scheme where one neural network is learned and shared by all agents in the environment. We show that our framework is both scalable and efficient for large number of incoming aircraft to achieve extremely high traffic throughput with safety guarantee. We evaluate our model via extensive simulations in the BlueSky environment. Results show that our framework is able to resolve 99.97% and 100% of all conflicts both at intersections and merging points, respectively, in extreme high-density air traffic scenarios.

Training agents with reinforcement learning based techniques requires thousands of steps, which translates to long training periods when applied to robots. By training the policy in a simulated environment we avoid such limitation. Typically, the action spaces in a simulation and real robot are kept as similar as possible, but if we want to use a generic simulation environment, this strategy will not work. Video games, such as Doom (1993), offer a crude but multi-purpose environments that can used for learning various tasks. However, original Doom has four discrete actions for movement and the robot in our case has two continuous actions. In this work, we study the transfer between these two different action spaces. We begin with experiments in a simulated environment, after which we validate the results with experiments on a real robot. Results show that fine-tuning initially learned network parameters leads to unreliable results, but by keeping most of the neural network frozen we obtain above $90\%$ success rate in simulation and real robot experiments.

Reinforcement learning (RL) is about sequential decision making and is traditionally opposed to supervised learning (SL) and unsupervised learning (USL). In RL, given the current state, the agent makes a decision that may influence the next state as opposed to SL (and USL) where, the next state remains the same, regardless of the decisions taken, either in batch or online learning. Although this difference is fundamental between SL and RL, there are connections that have been overlooked. In particular, we prove in this paper that gradient policy method can be cast as a supervised learning problem where true label are replaced with discounted rewards. We provide a new proof of policy gradient methods (PGM) that emphasizes the tight link with the cross entropy and supervised learning. We provide a simple experiment where we interchange label and pseudo rewards. We conclude that other relationships with SL could be made if we modify the reward functions wisely.