This course is all about the application of deep learning and neural networks to reinforcement learning. If you've taken my first reinforcement learning class, then you know that reinforcement learning is on the bleeding edge of what we can do with AI. Specifically, the combination of deep learning with reinforcement learning has led to AlphaGo beating a world champion in the strategy game Go, it has led to self-driving cars, and it has led to machines that can play video games at a superhuman level. Reinforcement learning has been around since the 70s but none of this has been possible until now. The world is changing at a very fast pace. The state of California is changing their regulations so that self-driving car companies can test their cars without a human in the car to supervise.

The efficient use of limited computational resources is an essential ingredient of intelligence. Selecting computations optimally according to rational metareasoning would achieve this, but this is computationally intractable. Inspired by psychology and neuroscience, we propose the first concrete and domain-general learning algorithm for approximating the optimal selection of computations: Bayesian metalevel policy search (BMPS). We derive this general, sample-efficient search algorithm for a computation-selecting metalevel policy based on the insight that the value of information lies between the myopic value of information and the value of perfect information. We evaluate BMPS on three increasingly difficult metareasoning problems: when to terminate computation, how to allocate computation between competing options, and planning. Across all three domains, BMPS achieved near-optimal performance and compared favorably to previously proposed metareasoning heuristics. Finally, we demonstrate the practical utility of BMPS in an emergency management scenario, even accounting for the overhead of metareasoning.

Time travel and unchaining the time-matter continuum is no big deal. Nor is recruiting a dragon slayer, a Jedi Knight and a Transformer – a child's mind is able to create fantastic worlds in seconds. So what would happen if robots had an artificial imagination? Developing innovative strategies in Go or unorthodox approaches to chess are just top-of-mind examples of how the agent in reinforcement learning can be creative. Go, Chess and League of Legends all draw on the imagination: players use abstract thinking to predict their opponent's actions and construct a strategy for upcoming moves.

A team of researchers in the UK have taught an autonomous car to stay in its own damn lane in just 20 minutes -- an impressive feat considering I know a few human drivers that couldn't achieve that in their lifetime. Road rage aside, the team at Wayve, a company founded by researchers from Cambridge University's Engineering Department detailed their "reinforcement learning" algorithm in a blog post on June 28. The algorithm, in tandem with a human safety driver, taught the car how to remain within a lane over a period of "15-20 minutes." Reinforcement learning for AI has been shown to be highly effective before, with DeepMind Technologies showing it can learn how to play games such as Go or Chess and OpenAI showing that its AI plays 180 days worth of Dota 2 every single day. While defeating human players in incredible complex games like Go or Dota 2 is certainly impressive, teaching a car to drive itself is another wheelhouse altogether.

The recently proposed distributional approach to reinforcement learning (DiRL) is centered on learning the distribution of the reward-to-go, often referred to as the value distribution. In this work, we show that the distributional Bellman equation, which drives DiRL methods, is equivalent to a generative adversarial network (GAN) model. In this formulation, DiRL can be seen as learning a deep generative model of the value distribution, driven by the discrepancy between the distribution of the current value, and the distribution of the sum of current reward and next value. We use this insight to propose a GAN-based approach to DiRL, which leverages the strengths of GANs in learning distributions of high-dimensional data. In particular, we show that our GAN approach can be used for DiRL with multivariate rewards, an important setting which cannot be tackled with prior methods. The multivariate setting also allows us to unify learning the distribution of values and state transitions, and we exploit this idea to devise a novel exploration method that is driven by the discrepancy in estimating both values and states.

Urban Traffic Control (UTC) plays an essential role in Intelligent Transportation System (ITS) but remains difficult. Since model-based UTC methods may not accurately describe the complex nature of traffic dynamics in all situations, model-free data-driven UTC methods, especially reinforcement learning (RL) based UTC methods, received increasing interests in the last decade. However, existing DL approaches did not propose an efficient algorithm to solve the complicated multiple intersections control problems whose state-action spaces are vast. To solve this problem, we propose a Deep Reinforcement Learning (DRL) algorithm that combines several tricks to master an appropriate control strategy within an acceptable time. This new algorithm relaxes the fixed traffic demand pattern assumption and reduces human invention in parameter tuning. Simulation experiments have shown that our method outperforms traditional rule-based approaches and has the potential to handle more complex traffic problems in the real world.

We give a simple optimistic algorithm for which it is easy to derive regret bounds of $\tilde{O}(\sqrt{t_{\rm mix} SAT})$ after $T$ steps in uniformly ergodic MDPs with $S$ states, $A$ actions, and mixing time parameter $t_{\rm mix}$. These bounds are the first regret bounds in the general, non-episodic setting with an optimal dependence on all given parameters. They could only be improved by using an alternative mixing time parameter.

Learning to cooperate with friends and compete with foes is a key component of multi-agent reinforcement learning. Typically to do so, one requires access to either a model of or interaction with the other agent(s). Here we show how to learn effective strategies for cooperation and competition in an asymmetric information game with no such model or interaction. Our approach is to encourage an agent to reveal or hide their intentions using an information-theoretic regularizer. We consider both the mutual information between goal and action given state, as well as the mutual information between goal and state. We show how to stochastically optimize these regularizers in a way that is easy to integrate with policy gradient reinforcement learning. Finally, we demonstrate that cooperative (competitive) policies learned with our approach lead to more (less) reward for a second agent in two simple asymmetric information games.

In Reinforcement Learning, the agents take random decisions in their environment and learns on selecting the right one out of many to achieve their goal and play at a super-human level. Policy and Value Networks are used together in algorithms like Monte Carlo Tree Search to perform Reinforcement Learning. Both the networks are an integral part of a method called Exploration in MCTS algorithm. They are also known as policy iteration & value iteration since they are calculated many times making it an iterative process. Let's understand why are they so important in Machine Learning and what's the difference between them?