Documentation is available online: https://rllab.readthedocs.org/en/latest/.

This project follows the description of the Deep Q Learning algorithm described in Playing Atari with Deep Reinforcement Learning [2] and shows that this learning algorithm can be further generalized to the notorious Flappy Bird. It is a convolutional neural network, trained with a variant of Q-learning, whose input is raw pixels and whose output is a value function estimating future rewards. Since deep Q-network is trained on the raw pixel values observed from the game screen at each time step, [3] finds that remove the background appeared in the original game can make it converge faster. The architecture of the network is shown in the figure below. The first layer convolves the input image with an 8x8x4x32 kernel at a stride size of 4. The output is then put through a 2x2 max pooling layer.

The DeepFund Agent will make trading decisions directly from raw market data using Deep Learning, Deep Reinforcement Learning and Unsupervised Learning. The Agent was ordered to maximize the value of our bank account... The DeepFund Agent learns to trade from its experience and improves itself to a superhuman level.

Though substantial research has been dedicated towards using technology to improve education, no current methods are as effective as one-on-one tutoring. A critical, though relatively understudied, aspect of effective tutoring is modulating the student's affective state throughout the tutoring session in order to maximize long-term learning gains. We developed an integrated experimental paradigm in which children play a second-language learning game on a tablet, in collaboration with a fully autonomous social robotic learning companion. As part of the system, we measured children's valence and engagement via an automatic facial expression analysis system. These signals were combined into a reward signal that fed into the robot's affective reinforcement learning algorithm. Over several sessions, the robot played the game and personalized its motivational strategies (using verbal and non-verbal actions) to each student. We evaluated this system with 34 children in preschool classrooms for a duration of two months. We saw that (1) children learned new words from the repeated tutoring sessions, (2) the affective policy personalized to students over the duration of the study, and (3) students who interacted with a robot that personalized its affective feedback strategy showed a significant increase in valence, as compared to students who interacted with a non-personalizing robot. This integrated system of tablet-based educational content, affective sensing, affective policy learning, and an autonomous social robot holds great promise for a more comprehensive approach to personalized tutoring.

A construct that has been receiving attention recently in reinforcement learning is stochastic factorization (SF), a particular case of non-negative factorization (NMF) in which the matrices involved are stochastic. The idea is to use SF to approximate the transition matrices of a Markov decision process (MDP). This is useful for two reasons. First, learning the factors of the SF instead of the transition matrices can reduce significantly the number of parameters to be estimated. Second, it has been shown that SF can be used to reduce the number of operations needed to compute an MDP's value function. Recently, an algorithm called expectation-maximization SF (EMSF) has been proposed to compute a SF directly from transitions sampled from an MDP. In this paper we take a closer look at EMSF. First, by exploiting the assumptions underlying the algorithm, we show that it is possible to reduce it to simple multiplicative update rules similar to the ones that helped popularize NMF. Second, we analyze the optimization process underlying EMSF and find that it minimizes a modified version of the Kullback-Leibler divergence that is particularly well-suited for learning a SF from data sampled from an arbitrary distribution. Third, we build on this improved understanding of EMSF to draw an interesting connection with NMF and probabilistic latent semantic analysis. We also exploit the simplified update rules to introduce a new version of EMSF that generalizes and significantly improves its precursor. This new algorithm provides a practical mechanism to control the trade-off between memory usage and computing time, essentially freeing the space complexity of EMSF from its dependency on the number of sample transitions. The algorithm can also compute its approximation incrementally, which makes it possible to use it concomitantly with the collection of data. This feature makes the new version of EMSF particularly suitable for online reinforcement learning. Empirical results support the utility of the proposed algorithm.

Large state spaces and the curse of dimensionality contribute to the complexity of a task. Learning from demonstration techniques can be combined with reinforcement learning to narrow the exploration space of an agent, but require consistent and accurate demonstrations, as well as the state-action pairs for an entire demonstration. Individuals with severe motor disabilities are often slow and prone to human errors in demonstrations while teaching. My dissertation develops tools to allow persons with severe motor disabilities, and individuals in general, to train these systems. To handle these large state spaces as well as human error, we developed Dimensionality Reduced Reinforcement Learning. To accommodate slower feedback, we will develop a movie-reel style learning from demonstration interface.

Many planning methods rely on the use of an immediate reward function as a portable and succinct representation of desired behavior. Rewards are often inferred from demonstrated behavior that is assumed to be near-optimal. We examine a framework, Distance Minimization IRL (DM-IRL), for learning reward functions from scores an expert assigns to possibly suboptimal demonstrations. By changing the expert’s role from a demonstrator to a judge, DM-IRL relaxes some of the assumptions present in IRL, enabling learning from the scoring of arbitrary demonstration trajectories with unknown transition functions. DM-IRL complements existing IRL approaches by addressing different assumptions about the expert. We show that DM-IRL is robust to expert scoring error and prove that finding a policy that produces maximally informative trajectories for an expert to score is strongly NP-hard. Experimentally, we demonstrate that the reward function DM-IRL learns from an MDP with an unknown transition model can transfer to an agent with known characteristics in a novel environment, and we achieve successful learning with limited available training data.

We propose a new class of computationally fast algorithms to find close to optimal policy for Markov Decision Processes (MDP) with large finite horizon T.The main idea is that instead of planning until the time horizon T, we plan only up to a truncated horizon H << T and use an estimate of the true optimal value function as the terminal value. Our approach of finding the terminal value function is to learn a mapping from an MDP to its value function by solving many similar MDPs during a training phase and fit a regression estimator. We analyze the method by providing an error propagation theorem that shows the effect of various sources of errors to the quality of the solution. We also empirically validate this approach in a real-world application of designing an energy management system for Hybrid Electric Vehicles with promising results.

There are two classes of average reward reinforcement learning (RL) algorithms: model-based ones that explicitly maintain MDP models and model-free ones that do not learn such models. Though model-free algorithms are known to be more efficient, they often cannot converge to optimal policies due to the perturbation of parameters. In this paper, a novel model-free algorithm is proposed, which makes use of constant shifting values (CSVs) estimated from prior knowledge. To encourage exploration during the learning process, the algorithm constantly subtracts the CSV from the rewards. A terminating condition is proposed to handle the unboundedness of Q-values caused by such substraction. The convergence of the proposed algorithm is proved under very mild assumptions. Furthermore, linear function approximation is investigated to generalize our method to handle large-scale tasks. Extensive experiments on representative MDPs and the popular game Tetris show that the proposed algorithms significantly outperform the state-of-the-art ones.

Specifying a numeric reward function for reinforcement learning typically requires a lot of hand-tuning from a human expert. In contrast, preference-based reinforcement learning (PBRL) utilizes only pairwise comparisons between trajectories as a feedback signal, which are often more intuitive to specify. Currently available approaches to PBRL for control problems with continuous state/action spaces require a known or estimated model, which is often not available and hard to learn. In this paper, we integrate preference-based estimation of the reward function into a model-free reinforcement learning (RL) algorithm, resulting in a model-free PBRL algorithm. Our new algorithm is based on Relative Entropy Policy Search (REPS), enabling us to utilize stochastic policies and to directly control the greediness of the policy update. REPS decreases exploration of the policy slowly by limiting the relative entropy of the policy update, which ensures that the algorithm is provided with a versatile set of trajectories, and consequently with informative preferences. The preference-based estimation is computed using a sample-based Bayesian method, which can also estimate the uncertainty of the utility. Additionally, we also compare to a linear solvable approximation, based on inverse RL. We show that both approaches perform favourably to the current state-of-the-art. The overall result is an algorithm that can learn non-parametric continuous action policies from a small number of preferences.