Automatic continuous speech recognition (CSR) is sufficiently mature that a variety of real world applications are now possible including large vocabulary transcription and interactive spoken dialogues. This paper reviews the evolution of the statistical modelling techniques which underlie current-day systems, specifically hidden Markov models (HMMs) and N-grams. Starting from a description of the speech signal and its parameterisation, the various modelling assumptions and their consequences are discussed. It then describes various techniques by which the effects of these assumptions can be mitigated. Despite the progress that has been made, the limitations of current modelling techniques are still evident. The paper therefore concludes with a brief review of some of the more fundamental modelling work now in progress.

There exist a number of reinforcement learning algorithms which learnby climbing the gradient of expected reward. Their long-runconvergence has been proved, even in partially observableenvironments with non-deterministic actions, and without the need fora system model. However, the variance of the gradient estimator hasbeen found to be a significant practical problem. Recent approacheshave discounted future rewards, introducing a bias-variance trade-offinto the gradient estimate. We incorporate a reward baseline into thelearning system, and show that it affects variance without introducingfurther bias. In particular, as we approach the zero-bias,high-variance parameterization, the optimal (or variance minimizing)constant reward baseline is equal to the long-term average expectedreward. Modified policy-gradient algorithms are presented, and anumber of experiments demonstrate their improvement over previous work.

With the advance of efficient analytical methods for sensitivity analysis ofprobabilistic networks, the interest in the sensitivities revealed by real-life networks is rekindled. As the amount of data resulting from a sensitivity analysis of even a moderately-sized network is alreadyoverwhelming, methods for extracting relevant information are called for. One such methodis to study the derivative of the sensitivity functions yielded for a network's parameters. We further propose to build upon the concept of admissible deviation, that is, the extent to which a parameter can deviate from the true value without inducing a change in the most likely outcome. We illustrate these concepts by means of a sensitivity analysis of a real-life probabilistic network in oncology.

A Bayesian Belief Network (BN) is a model of a joint distribution over a setof n variables, with a DAG structure to represent the immediate dependenciesbetween the variables, and a set of parameters (aka CPTables) to represent thelocal conditional probabilities of a node, given each assignment to itsparents. In many situations, these parameters are themselves random variables - this may reflect the uncertainty of the domain expert, or may come from atraining sample used to estimate the parameter values. The distribution overthese "CPtable variables" induces a distribution over the response the BNwill return to any "What is Pr(H | E)?" query. This paper investigates thevariance of this response, showing first that it is asymptotically normal,then providing its mean and asymptotical variance. We then present aneffective general algorithm for computing this variance, which has the samecomplexity as simply computing the (mean value of) the response itself - ie,O(n 2^w), where n is the number of variables and w is the effective treewidth. Finally, we provide empirical evidence that this algorithm, whichincorporates assumptions and approximations, works effectively in practice,given only small samples.

We propose a new method of discovering causal structures, based on the detection of local, spontaneous changes in the underlying data-generating model. We analyze the classes of structures that are equivalent relative to a stream of distributions produced by local changes, and devise algorithms that output graphical representations of these equivalence classes. We present experimental results, using simulated data, and examine the errors associated with detection of changes and recovery of structures.

Chow and Liu (1968) studied the problem of learning a maximumlikelihood Markov tree. We generalize their work to more complexMarkov networks by considering the problem of learning a maximumlikelihood Markov network of bounded complexity. We discuss howtree-width is in many ways the appropriate measure of complexity andthus analyze the problem of learning a maximum likelihood Markovnetwork of bounded tree-width.Similar to the work of Chow and Liu, we are able to formalize thelearning problem as a combinatorial optimization problem on graphs. Weshow that learning a maximum likelihood Markov network of boundedtree-width is equivalent to finding a maximum weight hypertree. Thisequivalence gives rise to global, integer-programming based,approximation algorithms with provable performance guarantees, for thelearning problem. This contrasts with heuristic local-searchalgorithms which were previously suggested (e.g. by Malvestuto 1991).The equivalence also allows us to study the computational hardness ofthe learning problem. We show that learning a maximum likelihoodMarkov network of bounded tree-width is NP-hard, and discuss thehardness of approximation.

We present a new method for estimating the expected return of a POMDP from experience. The estimator does not assume any knowledge of the POMDP, can estimate the returns for finite state controllers, allows experience to be gathered from arbitrary sequences of policies, and estimates the return for any new policy. We motivate the estimator from function-approximation and importance sampling points-of-view and derive its bias and variance. Although the estimator is biased, it has low variance and the bias is often irrelevant when the estimator is used for pairwise comparisons. We conclude by extending the estimator to policies with memory and compare its performance in a greedy search algorithm to the REINFORCE algorithm showing an order of magnitude reduction in the number of trials required.

We consider a partially observable Markov decision problem (POMDP) that models a class of sequencing problems. Although POMDPs are typically intractable, our formulation admits tractable solution. Instead of maintaining a value function over a high-dimensional set of belief states, we reduce the state space to one of smaller dimension, in which grid-based dynamic programming techniques are effective. We develop an error bound for the resulting approximation, and discuss an application of the model to a problem in targeted advertising.

We investigate a model for planning under uncertainty with temporallyextended actions, where multiple actions can be taken concurrently at each decision epoch. Our model is based on the options framework, and combines it with factored state space models,where the set of options can be partitioned into classes that affectdisjoint state variables. We show that the set of decisionepochs for concurrent options defines a semi-Markov decisionprocess, if the underlying temporally extended actions being parallelized arerestricted to Markov options. This property allows us to use SMDPalgorithms for computing the value function over concurrentoptions. The concurrent options model allows overlapping execution ofoptions in order to achieve higher performance or in order to performa complex task. We describe a simple experiment using a navigationtask which illustrates how concurrent options results in a faster planwhen compared to the case when only one option is taken at a time.

A method is presented for the rhythmic parsing problem: Given a sequence of observed musical note onset times, we estimate the corresponding notated rhythm and tempo process. A graphical model is developed that represents the simultaneous evolution of tempo and rhythm and relates these hidden quantities to observations. The rhythm variables are discrete and the tempo and observation variables are continuous. We show how to compute the globally most likely configuration of the tempo and rhythm variables given an observation of note onset times. Preliminary experiments are presented on a small data set. A generalization to arbitrary conditional Gaussian distributions is outlined.