Matthieu Geist

Kalman temporal differences

Because reinforcement learning suffers from a lack of scalability, online value (and Q-) function approximation has received increasing interest this last decade. This contribution introduces a novel approximation scheme, namely the Kalman Temporal Differences (KTD) framework, that exhibits the following features: sample-efficiency, non-linear approximation, non-stationarity handling and uncertainty management. A first KTD-based algorithm is provided for deterministic Markov Decision Processes (MDP) which produces biased estimates in the case of stochastic transitions. Than the eXtended KTD framework (XKTD), solving stochastic MDP, is described. Convergence is analyzed for special cases for both deterministic and stochastic transitions. Related algorithms are experimented on classical benchmarks. They compare favorably to the state of the art while exhibiting the announced features.

Sample-efficient batch reinforcement learning for dialogue management optimization

Spoken Dialogue Systems (SDS) are systems which have the ability to interact with human beings using natural language as the medium of interaction. A dialogue policy plays a crucial role in determining the functioning of the dialogue management module. Handcrafting the dialogue policy is not always an option, considering the complexity of the dialogue task and the stochastic behavior of users. In recent years approaches based on Reinforcement Learning (RL) for policy optimization in dialogue management have been proved to be an efficient approach for dialogue policy optimization. Yet most of the conventional RL algorithms are data intensive and demand techniques such as user simulation. Doing so, additional modeling errors are likely to occur. This paper explores the possibility of using a set of approximate dynamic programming algorithms for policy optimization in SDS. Moreover, these algorithms are combined to a method for learning a sparse representation of the value function. Experimental results show that these algorithms when applied to dialogue management optimization are particularly sample efficient, since they learn from few hundreds of dialogue examples. These algorithms learn in an off-policy manner, meaning that they can learn optimal policies with dialogue examples generated with a quite simple strategy. Thus they can learn good dialogue policies directly from data, avoiding user modeling errors.

Sample efficient on-line learning of optimal dialogue policies with kalman temporal differences

Designing dialog policies for voice-enabled interfaces is a tailoring job that is most often left to natural language processing experts. This job is generally redone for every new dialog task because cross-domain transfer is not possible. For this reason, machine learning methods for dialog policy optimization have been investigated during the last 15 years. Especially, reinforcement learning (RL) is now part of the state of the art in this domain. Standard RL methods require to test more or less random changes in the policy on users to assess them as improvements or degradations. This is called on policy learning. Nevertheless, it can result in system behaviors that are not acceptable by users. Learning algorithms should ideally infer an optimal strategy by observing interactions generated by a non-optimal but acceptable strategy, that is learning off-policy. In this contribution, a sample-efficient, online and off-policy reinforcement learning algorithm is proposed to learn an optimal policy from few hundreds of dialogues generated with a very simple handcrafted policy.

Algorithmic Survey of Parametric Value Function Approximation

Reinforcement learning (RL) is a machine learning answer to the optimal control problem. It consists of learning an optimal control policy through interactions with the system to be controlled, the quality of this policy being quantified by the so-called value function. A recurrent subtopic of RL concerns computing an approximation of this value function when the system is too large for an exact representation. This survey reviews state-of-the-art methods for (parametric) value function approximation by grouping them into three main categories: bootstrapping, residual, and projected fixed-point approaches. Related algorithms are derived by considering one of the associated cost functions and a specific minimization method, generally a stochastic gradient descent or a recursive least-squares approach.

Kalman Temporal Differences: The deterministic case

This paper deals with value function and Q-function approximation in deterministic Markovian decision processes. A general statistical framework based on the Kalman filtering paradigm is introduced. Its principle is to adopt a parametric representation of the value function, to model the associated parameter vector as a random variable and to minimize the mean-squared error of the parameters conditioned on past observed transitions. From this general framework, which will be called Kalman Temporal Differences (KTD), and using an approximation scheme called the unscented transform, a family of algorithms is derived, namely KTD-V, KTD-SARSA and KTD-Q, which aim respectively at estimating the value function of a given policy, the Q-function of a given policy and the optimal Q-function. The proposed approach holds for linear and nonlinear parameterization. This framework is discussed and potential advantages and shortcomings are highlighted.

ℓ 1 -Penalized projected bellman residual

We consider the task of feature selection for value function approximation in reinforcement learning. A promising approach consists in combining the Least-Squares Temporal Difference (LSTD) algorithm with l1 -regularization, which has proven to be effective in the supervised learning community. This has been done recently whit the LARS-TD algorithm, which replaces the projection operator of LSTD with an l1 -penalized projection and solves the corresponding fixed-point problem. However, this approach is not guaranteed to be correct in the general off-policy setting. We take a different route by adding an l1 -penalty term to the projected Bellman residual, which requires weaker assumptions while offering a comparable performance. However, this comes at the cost of a higher computational complexity if only a part of the regularization path is computed. Nevertheless, our approach ends up to a supervised learning problem, which let envision easy extensions to other penalties.

Off-policy learning in large-scale POMDP-based dialogue systems

Reinforcement learning (RL) is now part of the state of the art in the domain of spoken dialogue systems (SDS) optimisation. Most performant RL methods, such as those based on Gaussian Processes, require to test small changes in the policy to assess them as improvements or degradations. This process is called on policy learning. Nevertheless, it can result in system behaviours that are not acceptable by users. Learning algorithms should ideally infer an optimal strategy by observing interactions generated by a non-optimal but acceptable strategy, that is learning off-policy. Such methods usually fail to scale up and are thus not suited for real-world systems. In this contribution, a sample-efficient, online and off-policy RL algorithm is proposed to learn an optimal policy. This algorithm is combined to a compact non-linear value function representation (namely a multi-layers perceptron) enabling to handle large scale systems.

Parametric value function approximation: A unified view

Reinforcement learning (RL) is a machine learning answer to the optimal control problem. It consists of learning an optimal control policy through interactions with the system to be controlled, the quality of this policy being quantified by the so-called value function. An important RL subtopic is to approximate this function when the system is too large for an exact representation. This survey reviews and unifies state of the art methods for parametric value function approximation by grouping them into three main categories: bootstrapping, residuals and projected fixed-point approaches. Related algorithms are derived by considering one of the associated cost functions and a specific way to minimize it, almost always a stochastic gradient descent or a recursive least-squares approach.

A Cascaded Supervised Learning Approach to Inverse Reinforcement Learning

This paper considers the Inverse Reinforcement Learning (IRL) problem, that is inferring a reward function for which a demonstrated expert policy is optimal. We propose to break the IRL problem down into two generic Supervised Learning steps: this is the Cascaded Supervised IRL (CSI) approach. A classification step that defines a score function is followed by a regression step providing a reward function. A theoretical analysis shows that the demonstrated expert policy is near-optimal for the computed reward function. Not needing to repeatedly solve a Markov Decision Process (MDP) and the ability to leverage existing techniques for classification and regression are two important advantages of the CSI approach. It is furthermore empirically demonstrated to compare positively to state-of-the-art approaches when using only transitions sampled according to the expert policy, up to the use of some heuristics. This is exemplified on two classical benchmarks (the mountain car problem and a highway driving simulator).

Learning from Demonstrations: Is It Worth Estimating a Reward Function?

This paper provides a comparative study between Inverse Reinforcement Learning (IRL) and Apprenticeship Learning (AL). IRL and AL are two frameworks, using Markov Decision Processes (MDP), which are used for the imitation learning problem where an agent tries to learn from demonstrations of an expert. In the AL framework, the agent tries to learn the expert policy whereas in the IRL framework, the agent tries to learn a reward which can explain the behavior of the expert. This reward is then optimized to imitate the expert. One can wonder if it is worth estimating such a reward, or if estimating a policy is sufficient. This quite natural question has not really been addressed in the literature right now. We provide partial answers, both from a theoretical and empirical point of view.