We make an important connection to existing results in econometrics to describe an alternative formulation of inverse reinforcement learning (IRL). In particular, we describe an algorithm using Conditional Choice Probabilities (CCP), which are maximum likelihood estimates of the policy estimated from expert demonstrations, to solve the IRL problem. Using the language of structural econometrics, we re-frame the optimal decision problem and introduce an alternative representation of value functions due to (Hotz and Miller 1993). In addition to presenting the theoretical connections that bridge the IRL literature between Economics and Robotics, the use of CCPs also has the practical benefit of reducing the computational cost of solving the IRL problem. Specifically, under the CCP representation, we show how one can avoid repeated calls to the dynamic programming subroutine typically used in IRL. We show via extensive experimentation on standard IRL benchmarks that CCP-IRL is able to outperform MaxEnt-IRL, with as much as a 5x speedup and without compromising on the quality of the recovered reward function.
ISAE-SUPAERO, Université de Toulouse（高等航空航天学院，图卢兹大学）；ENAC, Université de Toulouse（国立民用航空学院，图卢兹大学）
We study how Reinforcement Learning can be employed to optimally control parameters in evolutionary algorithms. We control the mutation probability of a (1+1) evolutionary algorithm on the OneMax function. This problem is modeled as a Markov Decision Process and solved with Value Iteration via the known transition probabilities. It is then solved via Q-Learning, a Reinforcement Learning algorithm, where the exact transition probabilities are not needed. This approach also allows previous expert or empirical knowledge to be included into learning. It opens new perspectives, both formally and co…
Istituto di Analisi dei Sistemi ed Informatica “A. Ruberti”, Consiglio Nazionale delle Ricerche（意大利国家研究委员会“A.鲁贝尔蒂”系统分析与信息研究所）；University of Catania（卡塔尼亚大学）；Politecnico di Torino（都灵理工大学）
We characterize the reachability probabilities in stochastic directed graphs by means of reinforcement learning methods. In particular, we show that the dynamics of the transition probabilities in a stochastic digraph can be modeled via a difference inclusion, which, in turn, can be interpreted as a Markov decision process. Using the latter framework, we offer a methodology to design reward functions to provide upper and lower bounds on the reachability probabilities of a set of nodes for stochastic digraphs. The effectiveness of the proposed technique is demonstrated by application to the dif…
Designing optimal reward functions has been desired but extremely difficult in reinforcement learning (RL). When it comes to modern complex tasks, sophisticated reward functions are widely used to simplify policy learning yet even a tiny adjustment on them is expensive to evaluate due to the drastically increasing cost of training. To this end, we propose a hindsight reward tweaking approach by designing a novel paradigm for deep reinforcement learning to model the influences of reward functions within a near-optimal space. We simply extend the input observation with a condition vector linearl…
Reinforcement learning (RL) algorithms typically start tabula rasa, without any prior knowledge of the environment, and without any prior skills. This however often leads to low sample efficiency, requiring a large amount of interaction with the environment. This is especially true in a lifelong learning setting, in which the agent needs to continually extend its capabilities. In this paper, we examine how a pre-trained task-independent language model can make a goal-conditional RL agent more sample efficient. We do this by facilitating transfer learning between different related tasks. We exp…
Conditional Advantage Estimation for Reinforcement Learning in Large Reasoning Models
Shanghai Jiao Tong University（上海交通大学）；Shanghai Artificial Intelligence Laboratory（上海人工智能实验室）；Fudan University（复旦大学）；Renmin University of China（中国人民大学）；The Chinese University of Hong Kong（香港中文大学）
Seoul National University（首尔国立大学）；Samsung Electronics（三星电子）
We tackle the blackbox issue of deep neural networks in the settings of reinforcement learning (RL) where neural agents learn towards maximizing reward gains in an uncontrollable way. Such learning approach is risky when the interacting environment includes an expanse of state space because it is then almost impossible to foresee all unwanted outcomes and penalize them with negative rewards beforehand. Unlike reverse analysis of learned neural features from previous works, our proposed method \nj{tackles the blackbox issue by encouraging} an RL policy network to learn interpretable latent feat…
An approach to reasoning with default rules where the proportion of exceptions, or more generally the probability of encountering an exception, can be at least roughly assessed is presented. It is based on local uncertainty propagation rules which provide the best bracketing of a conditional probability of interest from the knowledge of the bracketing of some other conditional probabilities. A procedure that uses two such propagation rules repeatedly is proposed in order to estimate any simple conditional probability of interest from the available knowledge. The iterative procedure, that does …
Machine learning pipeline potentially consists of several stages of operations like data preprocessing, feature engineering and machine learning model training. Each operation has a set of hyper-parameters, which can become irrelevant for the pipeline when the operation is not selected. This gives rise to a hierarchical conditional hyper-parameter space. To optimize this mixed continuous and discrete conditional hierarchical hyper-parameter space, we propose an efficient pipeline search and configuration algorithm which combines the power of Reinforcement Learning and Bayesian Optimization. Em…
We show an alternative way of representing a Bayesian belief network by sensitivities and probability distributions. This representation is equivalent to the traditional representation by conditional probabilities, but makes dependencies between nodes apparent and intuitively easy to understand. We also propose a QR matrix representation for the sensitivities and/or conditional probabilities which is more efficient, in both memory requirements and computational speed, than the traditional representation for computer-based implementations of probabilistic inference. We use sensitivities to show…
Elicitation of probabilities is one of the most laborious tasks in building decision-theoretic models, and one that has so far received only moderate attention in decision-theoretic systems. We propose a set of user interface tools for graphical probabilistic models, focusing on two aspects of probability elicitation: (1) navigation through conditional probability tables and (2) interactive graphical assessment of discrete probability distributions. We propose two new graphical views that aid navigation in very large conditional probability tables: the CPTree (Conditional Probability Tree) and…
In Moral, Campos (1991) and Cano, Moral, Verdegay-Lopez (1991) a new method of conditioning convex sets of probabilities has been proposed. The result of it is a convex set of non-necessarily normalized probability distributions. The normalizing factor of each probability distribution is interpreted as the possibility assigned to it by the conditioning information. From this, it is deduced that the natural value for the conditional probability of an event is a possibility distribution. The aim of this paper is to study methods of transforming this possibility distribution into a probability (o…
In this paper, starting from a generalized coherent (i.e. avoiding uniform loss) intervalvalued probability assessment on a finite family of conditional events, we construct conditional probabilities with quasi additive classes of conditioning events which are consistent with the given initial assessment. Quasi additivity assures coherence for the obtained conditional probabilities. In order to reach our goal we define a finite sequence of conditional probabilities by exploiting some theoretical results on g-coherence. In particular, we use solutions of a finite sequence of linear systems.
Algorithms for learning the conditional probabilities of Bayesian networks with hidden variables typically operate within a high-dimensional search space and yield only locally optimal solutions. One way of limiting the search space and avoiding local optima is to impose qualitative constraints that are based on background knowledge concerning the domain. We present a method for integrating formal statements of qualitative constraints into two learning algorithms, APN and EM. In our experiments with synthetic data, this method yielded networks that satisfied the constraints almost perfectly. T…
Results on approximate deduction in the context of the calculus of evidence of Dempster-Shafer and the theory of interval probabilities are reported. Approximate conditional knowledge about the truth of conditional propositions was assumed available and expressed as sets of possible values (actually numeric intervals) of conditional probabilities. Under different interpretations of this conditional knowledge, several formulas were produced to integrate unconditioned estimates (assumed given as sets of possible values of unconditioned probabilities) with conditional estimates. These formulas ar…
We present our submission to the SIGTYP 2020 Shared Task on the prediction of typological features. We submit a constrained system, predicting typological features only based on the WALS database. We investigate two approaches. The simpler of the two is a system based on estimating correlation of feature values within languages by computing conditional probabilities and mutual information. The second approach is to train a neural predictor operating on precomputed language embeddings based on WALS features. Our submitted system combines the two approaches based on their self-estimated confiden…
McDonnell Douglas Research Laboratories（麦克唐纳-道格拉斯研究实验室）
Expert systems applications that involve uncertain inference can be represented by a multidimensional contingency table. These tables offer a general approach to inferring with uncertain evidence, because they can embody any form of association between any number of pieces of evidence and conclusions. (Simpler models may be required, however, if the number of pieces of evidence bearing on a conclusion is large.) This paper presents a method of using these tables to make uncertain inferences without assumptions of conditional independence among pieces of evidence or heuristic combining rules. A…
Northeastern Illinois University（东北伊利诺伊大学）；Illinois Institute of Technology（伊利诺伊理工学院）
The belief network is a well-known graphical structure for representing independences in a joint probability distribution. The methods, which perform probabilistic inference in belief networks, often treat the conditional probabilities which are stored in the network as certain values. However, if one takes either a subjectivistic or a limiting frequency approach to probability, one can never be certain of probability values. An algorithm should not only be capable of reporting the probabilities of the alternatives of remaining nodes when other nodes are instantiated; it should also be capable…
Univ. of Michigan（密歇根大学）；US Army DEVCOM ARL（美国陆军作战能力发展司令部陆军研究实验室）；Univ. of Brescia（布雷西亚大学）
When the historical data are limited, the conditional probabilities associated with the nodes of Bayesian networks are uncertain and can be empirically estimated. Second order estimation methods provide a framework for both estimating the probabilities and quantifying the uncertainty in these estimates. We refer to these cases as uncer tain or second-order Bayesian networks. When such data are complete, i.e., all variable values are observed for each instantiation, the conditional probabilities are known to be Dirichlet-distributed. This paper improves the current state-of-the-art approaches f…
Joint Institute for Nuclear Research（联合核研究所）；Universidade de São Paulo（圣保罗大学）
An approach is presented treating decision theory as a probabilistic theory based on quantum techniques. Accurate definitions are given and thorough analysis is accomplished for the quantum probabilities describing the choice between separate alternatives, sequential alternatives characterizing conditional quantum probabilities, and behavioral quantum probabilities taking into account rational-irrational duality of decision making. The comparison between quantum and classical probabilities is explained. The analysis demonstrates that quantum probabilities serve as an essentially more powerful …
The problem is sequence prediction in the following setting. A sequence $x_1,...,x_n,...$ of discrete-valued observations is generated according to some unknown probabilistic law (measure) $μ$. After observing each outcome, it is required to give the conditional probabilities of the next observation. The measure $μ$ belongs to an arbitrary class $\C$ of stochastic processes. We are interested in predictors $ρ$ whose conditional probabilities converge to the "true" $μ$-conditional probabilities if any $μ\in\C$ is chosen to generate the data. We show that if such a predictor exists, then a predi…
