In this paper, we introduce a skill-balancing mechanism for adversarial non-player characters (NPCs), called Skilled Experience Catalogue (SEC). The objective of this mechanism is to approximately match the skill level of an NPC to an opponent in real-time. We test the technique in the context of a First-Person Shooter (FPS) game. Specifically, the technique adjusts a reinforcement learning NPC's proficiency with a weapon based on its current performance against an opponent. Firstly, a catalogue of experience, in the form of stored learning policies, is built up by playing a series of training games. Once the NPC has been sufficiently trained, the catalogue acts as a timeline of experience with incremental knowledge milestones in the form of stored learning policies. If the NPC is performing poorly, it can jump to a later stage in the learning timeline to be equipped with more informed decision-making. Likewise, if it is performing significantly better than the opponent, it will jump to an earlier stage. The NPC continues to learn in real-time using reinforcement learning but its policy is adjusted, as required, by loading the most suitable milestones for the current circumstances.

Neural networks allow Q-learning reinforcement learning agents such as deep Q-networks (DQN) to approximate complex mappings from state spaces to value functions. However, this also brings drawbacks when compared to other function approximators such as tile coding or their generalisations, radial basis functions (RBF) because they introduce instability due to the side effect of globalised updates present in neural networks. This instability does not even vanish in neural networks that do not have any hidden layers. In this paper, we show that simple modifications to the structure of the neural network can improve stability of DQN learning when a multi-layer perceptron is used for function approximation.

We address the problem of spatial segmentation of a 2D object in the context of a robotic system for painting, where an optimal segmentation depends on both the appearance of the object and the size of each segment. Since each segment must take into account appearance features at several scales, we take a hierarchical grammar-based parsing approach to decompose the object into 2D segments for painting. Since there are many ways to segment an object the solution space is extremely large and it is very challenging to utilize an exploration based optimization approach like reinforcement learning. Instead, we pose the segmentation problem as an imitation learning problem by using a segmentation algorithm in the place of an expert, that has access to a small dataset with known foreground-background segmentations. During the imitation learning process, we learn to imitate the oracle (segmentation algorithm) using only the image of the object, without the use of the known foreground-background segmentations. We introduce a novel deterministic policy gradient update, DRAG, in the form of a deterministic actor-critic variant of AggreVaTeD, to train our neural network based object parser. We will also show that our approach can be seen as extending DDPG to the Imitation Learning scenario. Training our neural parser to imitate the oracle via DRAG allow our neural parser to outperform several existing imitation learning approaches.

We propose a novel reinforcement learning approach for finite Markov decision processes (MDPs) with delayed rewards. In this work, biases of temporal difference (TD) estimates are proved to be corrected only exponentially slowly in the number of delay steps. Furthermore, variances of Monte Carlo (MC) estimates are proved to increase the variance of other estimates, the number of which can exponentially grow in the number of delay steps. We introduce RUDDER, a return decomposition method, which creates a new MDP with same optimal policies as the original MDP but with redistributed rewards that have largely reduced delays. If the return decomposition is optimal, then the new MDP does not have delayed rewards and TD estimates are unbiased. In this case, the rewards track Q-values so that the future expected reward is always zero. We experimentally confirm our theoretical results on bias and variance of TD and MC estimates. On artificial tasks with different lengths of reward delays, we show that RUDDER is exponentially faster than TD, MC, and MC Tree Search (MCTS). RUDDER outperforms rainbow, A3C, DDQN, Distributional DQN, Dueling DDQN, Noisy DQN, and Prioritized DDQN on the delayed reward Atari game Venture in only a fraction of the learning time. RUDDER considerably improves the state-of-the-art on the delayed reward Atari game Bowling in much less learning time. Source code is available at https://github.com/ml-jku/baselines-rudder, with demonstration videos at https://goo.gl/EQerZV.

We have seen much recent progress in rigid object manipulation, but interaction with deformable objects has notably lagged behind. Due to the large configuration space of deformable objects, solutions using traditional modelling approaches require significant engineering work. Perhaps then, bypassing the need for explicit modelling and instead learning the control in an end-to-end manner serves as a better approach? Despite the growing interest in the use of end-to-end robot learning approaches, only a small amount of work has focused on their applicability to deformable object manipulation. Moreover, due to the large amount of data needed to learn these end-to-end solutions, an emerging trend is to learn to control policies in simulation and then transfer them over to the real world. To-date, no work has explored whether it is possible to learn and transfer deformable object policies. We believe that if sim-to-real methods are the way forward, then it should be possible to learn to interact with a wide variety of objects, and not just rigid objects. In this work, we use a combination of state-of-the-art deep reinforcement learning algorithms to solve the problem of manipulating deformable objects (specifically cloth). We evaluate our approach on three tasks --- folding a towel up to a mark, folding a face towel diagonally, and draping a piece of cloth over a hanger. Our agents are fully trained in simulation with domain randomisation, and then successfully deployed in the real world without having seen any real deformable objects.

Our own Chief Data Scientist Dr. Tim Oates recently give a talk at the Thug Think conference in Portland, Oregon. He focused in on reinforcement learning, a branch of machine learning with exciting applications. Whether you're familiar or brand new to the subject, his engaging presentation is going to be of interest to you. If you want to hear more of Tim's thoughts on reinforcement learning, read "Exploration vs. Exploitation"

We introduce a novel method to learn a policy from unsupervised demonstrations of a process. Given a model of the system and a set of sequences of outputs, we find a policy that has a comparable performance to the original policy, without requiring access to the inputs of these demonstrations. We do so by first estimating the inputs of the system from observed unsupervised demonstrations. Then, we learn a policy by applying vanilla supervised learning algorithms to the (estimated)input-output pairs. For the input estimation, we present a new adaptive linear estimator (AdaL-IE) that explicitly trades-off variance and bias in the estimation. As we show empirically, AdaL-IE produces estimates with lower error compared to the state-of-the-art input estimation method, (UMV-IE) [Gillijns and De Moor, 2007]. Using AdaL-IE in conjunction with imitation learning enables us to successfully learn control policies that consistently outperform those using UMV-IE.

In this work we present a new agent architecture, called Reactor, which combines multiple algorithmic and architectural contributions to produce an agent with higher sample-efficiency than Prioritized Dueling DQN (Wang et al., 2016) and Categorical DQN (Bellemare et al., 2017), while giving better run-time performance than A3C (Mnih et al., 2016). Our first contribution is a new policy evaluation algorithm called Distributional Retrace, which brings multi-step off-policy updates to the distributional reinforcement learning setting. The same approach can be used to convert several classes of multi-step policy evaluation algorithms designed for expected value evaluation into distributional ones. Next, we introduce the \b{eta}-leave-one-out policy gradient algorithm which improves the trade-off between variance and bias by using action values as a baseline. Our final algorithmic contribution is a new prioritized replay algorithm for sequences, which exploits the temporal locality of neighboring observations for more efficient replay prioritization. Using the Atari 2600 benchmarks, we show that each of these innovations contribute to both the sample efficiency and final agent performance. Finally, we demonstrate that Reactor reaches state-of-the-art performance after 200 million frames and less than a day of training.

Unlike humans, who are capable of continual learning over their lifetimes, artificial neural networks have long been known to suffer from a phenomenon known as catastrophic forgetting, whereby new learning can lead to abrupt erasure of previously acquired knowledge. Whereas in a neural network the parameters are typically modelled as scalar values, an individual synapse in the brain comprises a complex network of interacting biochemical components that evolve at different timescales. In this paper, we show that by equipping tabular and deep reinforcement learning agents with a synaptic model that incorporates this biological complexity (Benna & Fusi, 2016), catastrophic forgetting can be mitigated at multiple timescales. In particular, we find that as well as enabling continual learning across sequential training of two simple tasks, it can also be used to overcome within-task forgetting by reducing the need for an experience replay database.