Advances in Reinforcement Learning (RL) have successfully tackled sample efficiency and overestimation bias. However, these methods often fall short of scalable performance. On the other hand, genetic methods provide scalability but depict hyperparameter sensitivity to evolutionary operations. We present the Evolution-based Soft Actor-Critic (ESAC), a scalable RL algorithm. Our contributions are threefold; ESAC (1) abstracts exploration from exploitation by combining Evolution Strategies (ES) with Soft Actor-Critic (SAC), (2) provides dominant skill transfer between offsprings by making use of soft winner selections and genetic crossovers in hindsight and (3) improves hyperparameter sensitivity in evolutions using Automatic Mutation Tuning (AMT). AMT gradually replaces the entropy framework of SAC allowing the population to succeed at the task while acting as randomly as possible, without making use of backpropagation updates. On a range of challenging control tasks consisting of high-dimensional action spaces and sparse rewards, ESAC demonstrates state-of-the-art performance and sample efficiency equivalent to SAC. ESAC demonstrates scalability comparable to ES on the basis of hardware resources and algorithm overhead. A complete implementation of ESAC with notes on reproducibility and videos can be found at the project website https://karush17.github.io/esac-web/.

A popular computer puzzle, the game of Minesweeper requires its human players to have a mix of both luck and strategy to succeed. Analyzing these aspects more formally, in our research we assessed the feasibility of a novel methodology based on Reinforcement Learning as an adequate approach to tackle the problem presented by this game. For this purpose we employed Multi-Armed Bandit algorithms which were carefully adapted in order to enable their use to define autonomous computational players, targeting to make the best use of some game peculiarities. After experimental evaluation, results showed that this approach was indeed successful, especially in smaller game boards, such as the standard beginner level. Despite this fact the main contribution of this work is a detailed examination of Minesweeper from a learning perspective, which led to various original insights which are thoroughly discussed.

Although duality is used extensively in certain fields, such as supervised learning in machine learning, it has been much less explored in others, such as reinforcement learning (RL). In this paper, we show how duality is involved in a variety of RL work, from that which spearheaded the field, such as Richard Bellman's value iteration, to that which was done within just the past few years yet has already had significant impact, such as TRPO, A3C, and GAIL. We show that duality is not uncommon in reinforcement learning, especially when value iteration, or dynamic programming, is used or when first or second order approximations are made to transform initially intractable problems into tractable convex programs.

Deep Q-learning algorithms often suffer from poor gradient estimations with an excessive variance, resulting in unstable training and poor sampling efficiency. Stochastic variance-reduced gradient methods such as SVRG have been applied to reduce the estimation variance (Zhao et al. 2019). However, due to the online instance generation nature of reinforcement learning, directly applying SVRG to deep Q-learning is facing the problem of the inaccurate estimation of the anchor points, which dramatically limits the potentials of SVRG. To address this issue and inspired by the recursive gradient variance reduction algorithm SARAH (Nguyen et al. 2017), this paper proposes to introduce the recursive framework for updating the stochastic gradient estimates in deep Q-learning, achieving a novel algorithm called SRG-DQN. Unlike the SVRG-based algorithms, SRG-DQN designs a recursive update of the stochastic gradient estimate. The parameter update is along an accumulated direction using the past stochastic gradient information, and therefore can get rid of the estimation of the full gradients as the anchors. Additionally, SRG-DQN involves the Adam process for further accelerating the training process. Theoretical analysis and the experimental results on well-known reinforcement learning tasks demonstrate the efficiency and effectiveness of the proposed SRG-DQN algorithm.

The combination of Monte-Carlo tree search (MCTS) with deep reinforcement learning has led to significant advances in artificial intelligence. However, AlphaZero, the current state-of-the-art MCTS algorithm, still relies on handcrafted heuristics that are only partially understood. In this paper, we show that AlphaZero's search heuristics, along with other common ones such as UCT, are an approximation to the solution of a specific regularized policy optimization problem. With this insight, we propose a variant of AlphaZero which uses the exact solution to this policy optimization problem, and show experimentally that it reliably outperforms the original algorithm in multiple domains.

Machine learning is a concept that is as old as computers. In 1950, Alan Turing created the Turning Test. It was a test for computers to see if a machine can convince a human it is a human and not a computer. Soon after that, in 1952, Arthur Samuel designed the first computer program where a computer can learn as it ran. This program was a checker game, where the computer learned the player's patterns during the match, and then use this knowledge to improve the computer's next moves.

Sequential decision making, commonly formalized as Markov Decision Process optimization, is a key challenge in artificial intelligence. Two successful approaches to MDP optimization are planning and reinforcement learning. Both research fields largely have their own research communities. However, if both research fields solve the same problem, then we should be able to disentangle the common factors in their solution approaches. Therefore, this paper presents a unifying framework for reinforcement learning and planning (FRAP), which identifies the underlying dimensions on which any planning or learning algorithm has to decide. At the end of the paper, we compare - in a single table - a variety of well-known planning, model-free and model-based RL algorithms along the dimensions of our framework, illustrating the validity of the framework. Altogether, FRAP provides deeper insight into the algorithmic space of planning and reinforcement learning, and also suggests new approaches to integration of both fields.

Sequential decision making, commonly formalized as Markov Decision Process (MDP) optimization, is a key challenge in artificial intelligence. Two key approaches to this problem are reinforcement learning (RL) and planning. This paper presents a survey of the integration of both fields, better known as model-based reinforcement learning. Model-based RL has two main steps. First, we systematically cover approaches to dynamics model learning, including challenges like dealing with stochasticity, uncertainty, partial observability, and temporal abstraction. Second, we present a systematic categorization of planning-learning integration, including aspects like: where to start planning, what budgets to allocate to planning and real data collection, how to plan, and how to integrate planning in the learning and acting loop. After these two key sections, we also discuss the potential benefits of model-based RL, like enhanced data efficiency, targeted exploration, and improved stability. Along the survey, we also draw connections to several related RL fields, like hierarchical RL and transfer, and other research disciplines, like behavioural psychology. Altogether, the survey presents a broad conceptual overview of planning-learning combinations for MDP optimization.

A fundamental challenge in reinforcement learning is to learn policies that generalize beyond the operating domains experienced during training. In this paper, we approach this challenge through the following invariance principle: an agent must find a representation such that there exists an action-predictor built on top of this representation that is simultaneously optimal across all training domains. Intuitively, the resulting invariant policy enhances generalization by finding causes of successful actions. We propose a novel learning algorithm, Invariant Policy Optimization (IPO), that implements this principle and learns an invariant policy during training. We compare our approach with standard policy gradient methods and demonstrate significant improvements in generalization performance on unseen domains for linear quadratic regulator and grid-world problems, and an example where a robot must learn to open doors with varying physical properties.

Designing task-oriented dialogue systems is a challenging research topic, since it needs not only to generate utterances fulfilling user requests but also to guarantee the comprehensibility. Many previous works trained end-to-end (E2E) models with supervised learning (SL), however, the bias in annotated system utterances remains as a bottleneck. Reinforcement learning (RL) deals with the problem through using non-differentiable evaluation metrics (e.g., the success rate) as rewards. Nonetheless, existing works with RL showed that the comprehensibility of generated system utterances could be corrupted when improving the performance on fulfilling user requests. In our work, we (1) propose modelling the hierarchical structure between dialogue policy and natural language generator (NLG) with the option framework, called HDNO; (2) train HDNO with hierarchical reinforcement learning (HRL), as well as suggest alternating updates between dialogue policy and NLG during HRL inspired by fictitious play, to preserve the comprehensibility of generated system utterances while improving fulfilling user requests; and (3) propose using a discriminator modelled with language models as an additional reward to further improve the comprehensibility. We test HDNO on MultiWoz 2.0 and MultiWoz 2.1, the datasets on multi-domain dialogues, in comparison with word-level E2E model trained with RL, LaRL and HDSA, showing a significant improvement on the total performance evaluated with automatic metrics.