Research in reinforcement learning has produced algorithms for optimal decision making under uncertainty that fall within two main types. The first employs a Bayesian framework, where optimality improves with increased computational time. This is because the resulting planning task takes the form of a dynamic programming problem on a belief tree with an infinite number of states. The second type employs relatively simple algorithm which are shown to suffer small regret within a distribution-free framework. This paper presents a lower bound and a high probability upper bound on the optimal value function for the nodes in the Bayesian belief tree, which are analogous to similar bounds in POMDPs. The bounds are then used to create more efficient strategies for exploring the tree. The resulting algorithms are compared with the distribution-free algorithm UCB1, as well as a simpler baseline algorithm on multi-armed bandit problems.
We state the problem of inverse reinforcement learning in terms of preference elicitation, resulting in a principled (Bayesian) statistical formulation. This generalises previous work on Bayesian inverse reinforcement learning and allows us to obtain a posterior distribution on the agent's preferences, policy and optionally, the obtained reward sequence, from observations. We examine the relation of the resulting approach to other statistical methods for inverse reinforcement learning via analysis and experimental results. We show that preferences can be determined accurately, even if the observed agent's policy is sub-optimal with respect to its own preferences. In that case, significantly improved policies with respect to the agent's preferences are obtained, compared to both other methods and to the performance of the demonstrated policy.
We present a class of models that, via a simple construction, enables exact, incremental, non-parametric, polynomial-time, Bayesian inference of conditional measures. The approach relies upon creating a sequence of covers on the conditioning variable and maintaining a different model for each set within a cover. Inference remains tractable by specifying the probabilistic model in terms of a random walk within the sequence of covers. We demonstrate the approach on problems of conditional density estimation, which, to our knowledge is the first closed-form, non-parametric Bayesian approach to this problem.
There has been a lot of recent work on Bayesian methods for reinforcement learning exhibiting near-optimal online performance. The main obstacle facing such methods is that in most problems of interest, the optimal solution involves planning in an infinitely large tree. However, it is possible to obtain stochastic lower and upper bounds on the value of each tree node. This enables us to use stochastic branch and bound algorithms to search the tree efficiently. This paper proposes two such algorithms and examines their complexity in this setting.
User authentication and intrusion detection differ from standard classification problems in that while we have data generated from legitimate users, impostor or intrusion data is scarce or non-existent. We review existing techniques for dealing with this problem and propose a novel alternative based on a principled statistical decision-making view point. We examine the technique on a toy problem and validate it on complex real-world data from an RFID based access control system. The results indicate that it can significantly outperform the classical world model approach. The method could be more generally useful in other decision-making scenarios where there is a lack of adversary data.
Several approximate policy iteration schemes without value functions, which focus on policy representation using classifiers and address policy learning as a supervised learning problem, have been proposed recently. Finding good policies with such methods requires not only an appropriate classifier, but also reliable examples of best actions, covering the state space sufficiently. Up to this time, little work has been done on appropriate covering schemes and on methods for reducing the sample complexity of such methods, especially in continuous state spaces. This paper focuses on the simplest possible covering scheme (a discretized grid over the state space) and performs a sample-complexity comparison between the simplest (and previously commonly used) rollout sampling allocation strategy, which allocates samples equally at each state under consideration, and an almost as simple method, which allocates samples only as needed and requires significantly fewer samples.
Intrusion Detection is an invaluable part of computer networks defense. An important consideration is the fact that raising false alarms carries a significantly lower cost than not detecting at- tacks. For this reason, we examine how cost-sensitive classification methods can be used in Intrusion Detection systems. The performance of the approach is evaluated under different experimental conditions, cost matrices and different classification models, in terms of expected cost, as well as detection and false alarm rates. We find that even under unfavourable conditions, cost-sensitive classification can improve performance significantly, if only slightly.
Several researchers have recently investigated the connection between reinforcement learning and classification. We are motivated by proposals of approximate policy iteration schemes without value functions which focus on policy representation using classifiers and address policy learning as a supervised learning problem. This paper proposes variants of an improved policy iteration scheme which addresses the core sampling problem in evaluating a policy through simulation as a multi-armed bandit machine. The resulting algorithm offers comparable performance to the previous algorithm achieved, however, with significantly less computational effort. An order of magnitude improvement is demonstrated experimentally in two standard reinforcement learning domains: inverted pendulum and mountain-car.
Supervised learning deals with the inference of a distribution over an output or label space $\CY$ conditioned on points in an observation space $\CX$, given a training dataset $D$ of pairs in $\CX \times \CY$. However, in a lot of applications of interest, acquisition of large amounts of observations is easy, while the process of generating labels is time-consuming or costly. One way to deal with this problem is {\em active} learning, where points to be labelled are selected with the aim of creating a model with better performance than that of an model trained on an equal number of randomly sampled points. In this paper, we instead propose to deal with the labelling cost directly: The learning goal is defined as the minimisation of a cost which is a function of the expected model performance and the total cost of the labels used. This allows the development of general strategies and specific algorithms for (a) optimal stopping, where the expected cost dictates whether label acquisition should continue (b) empirical evaluation, where the cost is used as a performance metric for a given combination of inference, stopping and sampling methods. Though the main focus of the paper is optimal stopping, we also aim to provide the background for further developments and discussion in the related field of active learning.