Learning about many things can provide numerous benefits to a reinforcement learning system. For example, learning many auxiliary value functions, in addition to optimizing the environmental reward, appears to improve both exploration and representation learning. The question we tackle in this paper is how to sculpt the stream of experience---how to adapt the system's behaviour---to optimize the learning of a collection of value functions. A simple answer is to compute an intrinsic reward based on the statistics of each auxiliary learner, and use reinforcement learning to maximize that intrinsic reward. Unfortunately, implementing this simple idea has proven difficult, and thus has been the focus of decades of study. It remains unclear which of the many possible measures of learning would work well in a parallel learning setting where environmental reward is extremely sparse or absent. In this paper, we investigate and compare different intrinsic reward mechanisms in a new bandit-like parallel-learning testbed. We discuss the interaction between reward and prediction learners and highlight the importance of introspective prediction learners: those that increase their rate of learning when progress is possible, and decrease when it is not. We provide a comprehensive empirical comparison of 15 different rewards, including well-known ideas from reinforcement learning and active learning. Our results highlight a simple but seemingly powerful principle: intrinsic rewards based on the amount of learning can generate useful behaviour, if each individual learner is introspective.

Efficient exploration is necessary to achieve good sample efficiency for reinforcement learning in general. From small, tabular settings such as gridworlds to large, continuous and sparse reward settings such as robotic object manipulation tasks, exploration through adding an uncertainty bonus to the reward function has been shown to be effective when the uncertainty is able to accurately drive exploration towards promising states. However reward bonuses can still be inefficient since they are non-stationary, which means that we must wait for function approximators to catch up and converge again when uncertainties change. We propose the idea of directed exploration, that is learning a goal-conditioned policy where goals are simply other states, and using that to directly try to reach states with large uncertainty. The goal-conditioned policy is independent of uncertainty and is thus stationary. We show in our experiments how directed exploration is more efficient at exploration and more robust to how the uncertainty is computed than adding bonuses to rewards.

Dyna is an architecture for model-based reinforcement learning (RL), where simulated experience from a model is used to update policies or value functions. A key component of Dyna is search-control, the mechanism to generate the state and action from which the agent queries the model, which remains largely unexplored. In this work, we propose to generate such states by using the trajectory obtained from Hill Climbing (HC) the current estimate of the value function. This has the effect of propagating value from high-value regions and of preemptively updating value estimates of the regions that the agent is likely to visit next. We derive a noisy stochastic projected gradient ascent algorithm for hill climbing, and highlight a connection to Langevin dynamics. We provide an empirical demonstration on four classical domains that our algorithm, HC-Dyna, can obtain significant sample efficiency improvements. We study the properties of different sampling distributions for search-control, and find that there appears to be a benefit specifically from using the samples generated by climbing on current value estimates from low-value to high-value region.

The performance of many algorithms in the fields of hard combinatorial problem solving, machine learning or AI in general depends on tuned hyperparameter configurations. Automated methods have been proposed to alleviate users from the tedious and error-prone task of manually searching for performance-optimized configurations across a set of problem instances. However there is still a lot of untapped potential through adjusting an algorithm's hyperparameters online since different hyperparameters are potentially optimal at different stages of the algorithm. We formulate the problem of adjusting an algorithm's hyperparameters for a given instance on the fly as a contextual MDP, making reinforcement learning (RL) the prime candidate to solve the resulting algorithm control problem in a data-driven way. Furthermore, inspired by applications of algorithm configuration, we introduce new white-box benchmarks suitable to study algorithm control. We show that on short sequences, algorithm configuration is a valid choice, but that with increasing sequence length a black-box view on the problem quickly becomes infeasible and RL performs better.

We provide a framework for incorporating robustness -- to perturbations in the transition dynamics which we refer to as model misspecification -- into continuous control Reinforcement Learning (RL) algorithms. We specifically focus on incorporating robustness into a state-of-the-art continuous control RL algorithm called Maximum a-posteriori Policy Optimization (MPO). We achieve this by learning a policy that optimizes for a worst case, entropy-regularized, expected return objective and derive a corresponding robust entropy-regularized Bellman contraction operator. In addition, we introduce a less conservative, soft-robust, entropy-regularized objective with a corresponding Bellman operator. We show that both, robust and soft-robust policies, outperform their non-robust counterparts in nine Mujoco domains with environment perturbations. Finally, we present multiple investigative experiments that provide a deeper insight into the robustness framework; including an adaptation to another continuous control RL algorithm as well as comparing this approach to domain randomization. Performance videos can be found online at https://sites.google.com/view/robust-rl.

We present a framework for learning to plan hierarchically in domains with unknown dynamics. We enhance planning performance by exploiting problem structure in several ways: (i) We simplify the search over plans by leveraging knowledge of skill objectives, (ii) Shorter plans are generated by enforcing aggressively hierarchical planning, (iii) We learn transition dynamics with sparse local models for better generalisation. Our framework decomposes transition dynamics into skill effects and success conditions, which allows fast planning by reasoning on effects, while learning conditions from interactions with the world. We propose a simple method for learning new abstract skills, using successful trajectories stemming from completing the goals of a curriculum. Learned skills are then refined to leverage other abstract skills and enhance subsequent planning. We show that both conditions and abstract skills can be learned simultaneously while planning, even in stochastic domains. Our method is validated in experiments of increasing complexity, with up to 2^100 states, showing superior planning to classic non-hierarchical planners or reinforcement learning methods. Applicability to real-world problems is demonstrated in a simulation-to-real transfer experiment on a robotic manipulator.

Intelligent agents need a physical understanding of the world to predict the impact of their actions in the future. While learning-based models of the environment dynamics have contributed to significant improvements in sample efficiency compared to model-free reinforcement learning algorithms, they typically fail to generalize to system states beyond the training data, while often grounding their predictions on non-interpretable latent variables. We introduce Interactive Differentiable Simulation (IDS), a differentiable physics engine, that allows for efficient, accurate inference of physical properties of rigid-body systems. Integrated into deep learning architectures, our model is able to accomplish system identification using visual input, leading to an interpretable model of the world whose parameters have physical meaning. We present experiments showing automatic task-based robot design and parameter estimation for nonlinear dynamical systems by automatically calculating gradients in IDS. When integrated into an adaptive model-predictive control algorithm, our approach exhibits orders of magnitude improvements in sample efficiency over model-free reinforcement learning algorithms on challenging nonlinear control domains.

A key challenge for Multiagent RL (Reinforcement Learning) is the design of agent-specific, local rewards that are aligned with sparse global objectives. In this paper, we introduce MERL (Multiagent Evolutionary RL), a hybrid algorithm that does not require an explicit alignment between local and global objectives. MERL uses fast, policy-gradient based learning for each agent by utilizing their dense local rewards. Concurrently, an evolutionary algorithm is used to recruit agents into a team by directly optimizing the sparser global objective. We explore problems that require coupling (a minimum number of agents required to coordinate for success), where the degree of coupling is not known to the agents. We demonstrate that MERL's integrated approach is more sample-efficient and retains performance better with increasing coupling orders compared to MADDPG, the state-of-the-art policy-gradient algorithm for multiagent coordination.

Solving complex, temporally-extended tasks is a long-standing problem in reinforcement learning (RL). We hypothesize that one critical element of solving such problems is the notion of compositionality. With the ability to learn concepts and sub-skills that can be composed to solve longer tasks, i.e. hierarchical RL, we can acquire temporally-extended behaviors. However, acquiring effective yet general abstractions for hierarchical RL is remarkably challenging. In this paper, we propose to use language as the abstraction, as it provides unique compositional structure, enabling fast learning and combinatorial generalization, while retaining tremendous flexibility, making it suitable for a variety of problems. Our approach learns an instruction-following low-level policy and a high-level policy that can reuse abstractions across tasks, in essence, permitting agents to reason using structured language. To study compositional task learning, we introduce an open-source object interaction environment built using the MuJoCo physics engine and the CLEVR engine. We find that, using our approach, agents can learn to solve to diverse, temporally-extended tasks such as object sorting and multi-object rearrangement, including from raw pixel observations. Our analysis find that the compositional nature of language is critical for learning diverse sub-skills and systematically generalizing to new sub-skills in comparison to non-compositional abstractions that use the same supervision.

We present a novel reinforcement learning-based natural media painting algorithm. Our goal is to reproduce a reference image using brush strokes and we encode the objective through observations. Our formulation takes into account that the distribution of the reward in the action space is sparse and training a reinforcement learning algorithm from scratch can be difficult. We present an approach that combines self-supervised learning and reinforcement learning to effectively transfer negative samples into positive ones and change the reward distribution. We demonstrate the benefits of our painting agent to reproduce reference images with brush strokes. The training phase takes about one hour and the runtime algorithm takes about 30 seconds on a GTX1080 GPU reproducing a 1000 800 image with 20,000 strokes.