Trainbot: a spoken dialog system using partially observable Markov decision processes

In contrast to traditional rule-based approaches to building spoken dialogue systems, recent research has shown that it is possible to implement all of the required functionality using statistical models trained using a combination of supervised learning and reinforcement learning. This approach to spoken dialogue is based on the mathematics of partially observable Markov decision processes (POMDPs) in which user inputs are treated as observations of some underlying belief state, and system responses are determined by a policy which maps belief states into actions. Virtually all current spoken dialogue systems are designed to operate in either a specific carefully defined domain such as restaurant information and appointment booking, or they have very limited conversational ability such as in Siri and Google Now. However, if voice is to become a significant input modality for accessing web-based information and services, then techniques will be needed to enable conversational spoken dialogue systems to operate within open domains. This talk will discuss methods by which current statistical approaches to spoken dialogue can be extended to cover much wider domains. It will be argued that unlike many other areas of machine learning, spoken dialogue systems always have a user on-hand to provide supervision. Hence spoken dialogue systems provide a unique opportunity to automatically adapt on large quantities of speech data without the need for costly annotation.

Over several years, we have developed an approach to spoken dialogue systems that includes rule-based and trainable dialogue managers, spoken language understanding and generation modules, and a comprehensive dialogue system architecture. We present a Reinforcement Learning-based dialogue system that goes beyond standard rule-based models and computes on-line decisions of the best dialogue moves. The key concept of this work is that we bridge the gap between manually written dialog models (e.g. rule-based) and adaptive computational models such as Partially Observable Markov Decision Processes (POMDP) based dialogue managers.

Probabilistic graphical models, such as partially observable Markov decision processes (POMDPs), have been used in stochastic spoken dialog systems to handle the inherent uncertainty in speech recognition and language understanding. Such dialog systems suffer from the fact that only a relatively small number of domain variables are allowed in the model, so as to ensure the generation of good-quality dialog policies. At the same time, the non-language perception modalities on robots, such as vision-based facial expression recognition and Lidar-based distance detection, can hardly be integrated into this process. In this paper, we use a probabilistic commonsense reasoner to “guide” our POMDP-based dialog manager, and present a principled, multimodal dialog management (MDM) framework that allows the robot's dialog belief state to be seamlessly updated by both observations of human spoken language, and exogenous events such as the change of human facial expressions. The MDM approach has been implemented and evaluated both in simulation and on a real mobile robot using guidance tasks.

A common problem in spoken dialogue systems is finding the intention of the user. This problem deals with obtaining one or several topics for each transcribed, possibly noisy, sentence of the user. In this work, we apply the recent unsupervised learning method, Hidden Topic Markov Models (HTMM), for finding the intention of the user in dialogues. This technique combines two methods of Latent Dirichlet Allocation (LDA) and Hidden Markov Model (HMM) in order to learn topics of documents. We show that HTMM can be also used for obtaining intentions for the noisy transcribed sentences of the user in spoken dialogue systems. We argue that in this way we can learn possible states in a speech domain which can be used in the design stage of its spoken dialogue system. Furthermore, we discuss that the learned model can be augmented and used in a POMDP (Partially Observable Markov Decision Process) dialogue manager of the spoken dialogue system.

The development of spoken dialogue systems is soon faced with the limitations of current speech recognition technology that make recognition errors a recurring problem for any dialogue system. Several studies have suggested that there is no clear correlation between speech recognition scores and user satisfaction, or the ability to complete the tasks underlying spoken dialogue [Yankelovich et al., 1995] [Dybkjaer et al., 1997], suggesting that a certain level of errors should not prevent spoken dialogue systems from being successful. However, most of the studies on speech recognition errors have concentrated either on parsing incomplete utterances or on global dialogue robustness, i.e. at task completion level [Allen et al., 1996] [Stromback and Jonsson, 1998] [Brandt-Pook et al., 1996]. There have been very few studies exploring the impact of speech recognition errors across the various component of a dialogue system, or specifically evaluating the impact of speech recognition errors on the overall dialogue behaviour.

The framework of partially observable Markov decision processes (POMDPs) offers a standard approach to model uncertainty in many robot tasks. Traditionally, POMDPs are formulated with optimality objectives. In this article, we study a different formulation of POMDPs with <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Boolean objectives</i> . For robotic domains that require a correctness guarantee of accomplishing tasks, Boolean objectives are natural formulations. We investigate the problem of POMDPs with a common Boolean objective: <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">safe reachability</i> , requiring that the robot eventually reaches a goal state with a probability above a threshold while keeping the probability of visiting unsafe states below a different threshold. Our approach builds upon the previous work that represents POMDPs with Boolean objectives using symbolic constraints. We employ a satisfiability modulo theories (SMTs) solver to efficiently search for solutions, i.e., policies or conditional plans that specify the action to take contingent on every possible event. A full policy or conditional plan is generally expensive to compute. To improve computational efficiency, we introduce the notion of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">partial conditional plans</i> that cover sampled events to approximate a full conditional plan. Our approach constructs a partial conditional plan parameterized by a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">replanning probability</i> . We prove that the failure rate of the constructed partial conditional plan is bounded by the replanning probability. Our approach allows users to specify an appropriate bound on the replanning probability to balance efficiency and correctness. Moreover, we update this bound properly to quickly detect whether the current partial conditional plan meets the bound and avoid unnecessary computation. In addition, to further improve the efficiency, we cache partial conditional plans for sampled belief states and reuse these cached plans if possible. We validate our approach in several robotic domains. The results show that our approach outperforms a previous policy synthesis approach for POMDPs with safe-reachability objectives in these domains. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Note to Practitioners</i> —This article was motivated by two observations. On the one hand, in robotics applications where uncertainty in sensing and actions is present, the solution to the classical partially observable Markov decision process (POMDP) formulation is expensive to compute in general. On the other hand, in certain practical scenarios, formulations other than the classical POMDP make a lot of sense and can provide flexibility in balancing efficiency and correctness. This article considers a modified POMDP formulation that includes a Boolean objective, namely safe reachability. This article uses the notion of a partial conditional plan. Rather than explicitly enumerating all possible observations to construct a full conditional plan, this work samples a subset of all observations to ensure bounded replanning probability. Our theoretical and empirical results show that the failure rate of the constructed partial conditional plan is bounded by the replanning probability. Moreover, these partial conditional plans can be cached to further improve the performance. Our results suggest that for domains where replanning is easy, increasing the replanning probability bound usually leads to better scalability, and for domains where replanning is difficult or impossible in some states, we can decrease the bound and allocate more computation time to achieve a higher success rate. Hence, in certain cases, the practitioner can take advantage of their knowledge of the problem domain to scale to larger problems. Preliminary physical experiments suggest that this approach is applicable to real-world robotic domains, but it requires a discrete representation of the workspace. How to deal with continuous workspace directly is an interesting future direction.

In this paper, we present a novel approach for dialog modeling, which extends the idea underlying the partially observable Markov Decision Processes (POMDPs), i. e. it allows for calculating the dialog policy in real-time and thereby increases the system flexibility. The use of statistical dialog models is particularly advantageous to react adequately to common errors of speech recognition systems. Comparing our results to the reference system (POMDP), we achieve a relative reduction of 31:6% of the average dialog length. Furthermore, the proposed system shows a relative enhancement of 64:4% of the sensitivity rate in the error recognition capabilities using the same specifity rate in both systems. The achieved results are based on the Air Travelling Information System with 21650 user utterances in 1585 natural spoken dialogs. Index Terms: dialog modeling, dialog strategy, spoken language understanding

&lt;p&gt;Partially Observable Markov Decision Processes (POMDPs) is a modeling framework that allow decision makers to make decisions in uncertain environments. This framework is particularly suited for sequential decision making problems for which the uncertainties are revealed in stages. Constrained POMDPs (CPOMDPs) extend this framework by also accounting for various restrictions in the environment. This PhD thesis focuses on POMDPs, CPOMDPs and its extensions, proposes novel solution methods and studies applications of POMDPs in disease screening. First, we focus on developing efficient versions of the well-known algorithms for POMDPs. POMDPs are notoriously difficult to solve optimally, and exact solution algorithms typically fail to solve the instances with more than a few core states. In this regard, we consider scalable algorithms solve the POMDP problems. Specifically, we propose distributed and parallel versions of the exact solution methods such as Monahan's algorithm and the incremental pruning algorithm as well as the heuristics such as Lovejoy's upper bound and lower bound methods. Next, we study the CPOMDPs and propose linear programming-based solution methods as a flexible approach for those. We perform a detailed numerical study with five different problem instances to showcase the viability of the proposed solution approach for CPOMDPs. Lastly, we propose multi-objective CPOMDP models for the breast cancer screening problem. These CPOMDP models extend the previously proposed POMDP models for breast cancer screening by considering different screening modalities and simultaneously optimizing for expected total quality-adjusted life years maximization and life-time risk minimization objectives.&lt;/p&gt;

Aiming to reduce energy consumption, minimize transmission latency, and optimize the long-term performance of Wireless Sensor Networks (WSN), this paper proposes an improved routing model of Directed Diffusion. We formulated the routing issue as a partially observable Markov decision process (POMDP), such that the optimal nodes selection policy yielded by maximizing the expected reward. Since the solution of POMDP is ready-made, the attention is concentrated only on the formulation of the routing issue into a POMDP model, more specifically, the definition of the tuple which contains states, actions, observations, transition probabilities, observation probabilities, and reward function. Our experimental study shows that the POMDP-based Directed Diffusion routing outperforms other routing techniques in terms of three widely used metrics including average dissipated energy, average delay, and average lifetime.

Motion planning with imperfect state information is a crucial capability for autonomous robots to operate reliably in uncertain and dynamic environments. Partially observable Markov decision processes (POMDPs) provide a principled general framework for planning under uncertainty. Using probabilistic sampling, point-based POMDP solvers have drastically improved the speed of POMDP planning, enabling us to handle moderately complex robotic tasks. However, robot motion planning tasks with long time horizons remains a severe obstacle for even the fastest point-based POMDP solvers today. This paper proposes Milestone Guided Sampling (MiGS), a new point-based POMDP solver, which exploits state space information to reduce effective planning horizons. MiGS samples a set of points, called milestones, from a robot’s state space and constructs a simplified representation of the state space from the sampled milestones. It then uses this representation of the state space to guide sampling in the belief space and tries to capture the essential features of the belief space with a small number of sampled points. Preliminary results are very promising. We tested MiGS in simulation on several difficult POMDPs that model distinct robotic tasks with long time horizons in both 2-D and 3-D environments. These POMDPs are impossible to solve with the fastest point-based solvers today, but MiGS solved them in a few minutes.

Resolving structural uncertainty in natural resources management using POMDP approaches

Partially observable Markov decision processes (POMDPs) are models for sequential decision-making under uncertainty and incomplete information. Machine learning methods typically train recurrent neural networks (RNN) as effective representations of POMDP policies that can efficiently process sequential data. However, it is hard to verify whether the POMDP driven by such RNN-based policies satisfies safety constraints, for instance, given by temporal logic specifications. We propose a novel method that combines techniques from machine learning with the field of formal methods: training an RNN-based policy and then automatically extracting a so-called finite-state controller (FSC) from the RNN. Such FSCs offer a convenient way to verify temporal logic constraints. Implemented on a POMDP, they induce a Markov chain, and probabilistic verification methods can efficiently check whether this induced Markov chain satisfies a temporal logic specification. Using such methods, if the Markov chain does not satisfy the specification, a byproduct of verification is diagnostic information about the states in the POMDP that are critical for the specification. The method exploits this diagnostic information to either adjust the complexity of the extracted FSC or improve the policy by performing focused retraining of the RNN. The method synthesizes policies that satisfy temporal logic specifications for POMDPs with up to millions of states, which are three orders of magnitude larger than comparable approaches.

Ambiguous partially observable Markov decision processes: Structural results and applications

In this article, we study the problem of controlling a partially observed Markov decision process (POMDP) to either aid or hinder the estimation of its state trajectory. We encode the estimation objectives via the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">smoother entropy</i> , which is the conditional entropy of the state trajectory given measurements and controls. Consideration of the smoother entropy contrasts with previous approaches that instead resort to marginal (or instantaneous) state entropies due to tractability concerns. By establishing novel expressions for the smoother entropy in terms of the POMDP belief state, we show that both the problems of minimizing and maximizing the smoother entropy in POMDPs can surprisingly be reformulated as belief-state Markov decision processes with concave cost and value functions. The significance of these reformulations is that they render the smoother entropy a tractable optimization objective, with structural properties amenable to the use of standard POMDP solution techniques for <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">both</i> active estimation and obfuscation. Simulations illustrate that optimization of the smoother entropy leads to superior trajectory estimation and obfuscation compared to alternative approaches.

POMDP (Partially Observable Markov Decision Process) is a mathematical framework that models planning under uncertainty. Solving a POMDP is an intractable problem and even the state of the art POMDP solvers are too computationally expensive for large domains. This is a major bottleneck. In this paper, we propose a new heuristic, called the pairwise heuristic, that can be used in a one-step greedy strategy to find a near optimal solution for POMDP problems very quickly. This approach is a good candidate for large problems where real-time solution is a necessity but exact optimality of the solution is not vital. The pairwise heuristic uses the optimal solutions for pairs of states. For each pair of states in the POMDP, we find the optimal sequence of actions to resolve the uncertainty and to maximize the reward, given that the agent is uncertain about which state of the pair it is in. Then we use these sequences as a heuristic and find the optimal action in each step of the greedy strategy using this heuristic. We have tested our method on the available large classical test benchmarks in various domains. The resulting total reward is close to, if not greater than, the total reward obtained by other state of the art POMDP solvers, while the time required to find the solution is always much less.