CauseOccam: Learning Interpretable Abstract Representations in Reinforcement Learning Environments via Model Sparsity

"I choose this restaurant because they have vegan sandwiches" could be a typical explanation we would expect from a human. However, current Reinforcement Learning (RL) techniques are not able to provide such explanations, when trained on raw pixels. RL algorithms for state-of-the-art benchmark environments are based on neural networks, which lack interpretability, because of the very factor that makes them so versatile – they have many parameters and intermediate representations. Enforcing safety guarantees is important when deploying RL agents in the real world, and guarantees require interpretability of the agent. Humans use short explanations that capture only the essential parts and often contain few causes to explain an effect. In our thesis, we address the problem of making RL agents understandable by humans. In addition to the safety concerns, the quest to mimic human-like reasoning is of general scientific interest, as it sheds light on the easy problem of consciousness. The problem of providing interpretable and simple causal explanations of agent's behavior is connected to the problem of learning good state representations. If we lack such a representation, any reasoning algorithm's outputs would be useless for interpretability, since even the "referents" of the "thoughts" of such a system would be obscure to us. One way to define simplicity of causal explanations via the sparsity of the Causal Model that describes the environment: the causal graph has the fewest edges connecting causes to their effects. For example, a model for choosing the restaurant that only depends on the cause "vegan" is simpler and more interpretable than a model that looks at each pixel of a photo of the menu of a restaurant, and possibly relies as well on spurious correlations, such the style of the menu. In this thesis, we propose a framework "CauseOccam" for model-based Reinforcement Learning where the model is regularized for simplicity in terms of sparsity of the causal graph it corresponds to. The framework contains a learned mapping from observations to latent features, and a model predicting latent features at the next time-steps given ones from the current time-step. The latent features are regularized with the sparsity of the model, compared to a more traditional regularization on the features themselves, or via a hand-crafted interpretability loss. To achieve sparsity, we use discrete Bernoulli variables with gradient estimation, and to find the best parameters, we use the primal-dual constrained formulation to achieve a target model quality. The novelty of this work is in learning jointly a sparse causal graph and the representation taking pixels as the input on RL environments. We test this framework on benchmark environments with non-trivial high-dimensional dynamics and show that it can uncover the causal graph with the fewest edges in the latent space. We describe the implications of our work for developing priors enforcing interpretability.

Learning Search Strategies from Human Demonstrations

Guillaume Pierre Luc De Chambrier

Decision making and planning with partial state information is a problem faced by all forms of intelligent entities. The formulation of a problem under partial state information leads to an exorbitant set of choices with associated probabilistic outcomes making its resolution difficult when using traditional planning methods. Human beings have acquired the ability of acting under uncertainty through education and self-learning. Transferring our know-how to artificial agents and robots will make it faster for them to learn and even improve upon us in tasks in which incomplete knowledge is available, which is the objective of this thesis. We model how humans reason with respect to their beliefs and transfer this knowledge in the form of a parameterised policy, following a Programming by Demonstration framework, to a robot apprentice for two spatial navigation tasks: the first task consists of localising a wooden block on a table and for the second task a power socket must be found and connected. In both tasks the human teacher and robot apprentice only rely on haptic and tactile information. We model the human and robot's beliefs by a probability density function which we update through recursive Bayesian state space estimation. To model the reasoning processes of human subjects performing the search tasks we learn a generative joint distribution over beliefs and actions (end-effector velocities) which were recorded during the executions of the task. For the first search task the direct mapping from belief to actions is learned whilst for the second task we incorporate a cost function used to adapt the policy parameters in a Reinforcement Learning framework and show a considerable improvement over solely learning the behaviour with respect to the distance taken to accomplish the task. Both search tasks above can be considered as active localisation as the uncertainty originates only from the position of the agent in the world. We consider searches in which both the position of the robot and features of the environment are uncertain. Given the unstructured nature of the belief a histogram parametrisation of the joint distribution of the robots position and features is necessary. However, naively doing so becomes quickly intractable as the space and time complexity is exponential. We demonstrate that by only parametrising the marginals and by memorising the parameters of the measurement likelihood functions we can recover the exact same solution as the naive parametrisations at a cost which is linear in space and time complexity.

EPFL2016

Machine learning for information retrieval

David Grangier

In this thesis, we explore the use of machine learning techniques for information retrieval. More specifically, we focus on ad-hoc retrieval, which is concerned with searching large corpora to identify the documents relevant to user queries. This identification is performed through a ranking task. Given a user query, an ad-hoc retrieval system ranks the corpus documents, so that the documents relevant to the query ideally appear above the others. In a machine learning framework, we are interested in proposing learning algorithms that can benefit from limited training data in order to identify a ranker likely to achieve high retrieval performance over unseen documents and queries. This problem presents novel challenges compared to traditional learning tasks, such as regression or classification. First, our task is a ranking problem, which means that the loss for a given query cannot be measured as a sum of an individual loss suffered for each corpus document. Second, most retrieval queries present a highly unbalanced setup, with a set of relevant documents accounting only for a very small fraction of the corpus. Third, ad-hoc retrieval corresponds to a kind of "double" generalization problem, since the learned model should not only generalize to new documents but also to new queries. Finally, our task also presents challenging efficiency constraints, since ad-hoc retrieval is typically applied to large corpora. The main objective of this thesis is to investigate the discriminative learning of ad-hoc retrieval models. For that purpose, we propose different models based on kernel machines or neural networks adapted to different retrieval contexts. The proposed approaches rely on different online learning algorithms that allow efficient learning over large corpora. The first part of the thesis focuses on text retrieval. In this case, we adopt a classical approach to the retrieval ranking problem, and order the text documents according to their estimated similarity to the text query. The assessment of semantic similarity between text items plays a key role in that setup and we propose a learning approach to identify an effective measure of text similarity. This identification is not performed relying on a set of queries with their corresponding relevant document sets, since such data are especially expensive to label and hence rare. Instead, we propose to rely on hyperlink data, since hyperlinks convey semantic proximity information that is relevant to similarity learning. This setup is hence a transfer learning setup, where we benefit from the proximity information encoded by hyperlinks to improve the performance over the ad-hoc retrieval task. We then investigate another retrieval problem, i.e. the retrieval of images from text queries. Our approach introduces a learning procedure optimizing a criterion related to the ranking performance. This criterion adapts our previous learning objective for learning textual similarity to the image retrieval problem. This yields an image ranking model that addresses the retrieval problem directly. This approach contrasts with previous research that relies on an intermediate image annotation task. Moreover, our learning procedure builds upon recent work on the online learning of kernel-based classifiers. This yields an efficient, scalable algorithm, which can benefit from recent kernels developed for image comparison. In the last part of the thesis, we show that the objective function used in the previous retrieval problems can be applied to the task of keyword spotting, i.e. the detection of given keywords in speech utterances. For that purpose, we formalize this problem as a ranking task: given a keyword, the keyword spotter should order the utterances so that the utterances containing the keyword appear above the others. Interestingly, this formulation yields an objective directly maximizing the area under the receiver operating curve, the most common keyword spotter evaluation measure. This objective is then used to train a model adapted to this intrinsically sequential problem. This model is then learned with a procedure derived from the algorithm previously introduced for the image retrieval task. To conclude, this thesis introduces machine learning approaches for ad-hoc retrieval. We propose learning models for various multi-modal retrieval setups, i.e. the retrieval of text documents from text queries, the retrieval of images from text queries and the retrieval of speech recordings from written keywords. Our approaches rely on discriminative learning and enjoy efficient training procedures, which yields effective and scalable models. In all cases, links with prior approaches were investigated and experimental comparisons were conducted.

EPFL2008

Using reinforcement learning for optimizing the reproduction of tasks in robot programming by demonstration

Florent Günter

As robots start pervading human environments, the need for new interfaces that would simplify human-robot interaction has become more pressing. Robot Programming by Demonstration (RbD) develops intuitive ways of programming robots, taking inspiration in strategies used by humans to transmit knowledge to apprentices. The user-friendliness of RbD is meant to allow lay users with no prior knowledge in computer science, electronics or mechanics to train robots to accomplish tasks the same way as they would with a co-worker. When a trainer teaches a task to a robot, he/she shows a particular way of fulfilling the task. For a robot to be able to learn from observing the trainer, it must be able to learn what the task entails (i.e. answer the so-called "What-to-imitate?" question), by inferring the user's intentions. But most importantly, the robot must be able to adapt its own controller to fit at best the demonstration (the so-called "How-to-imitate?" question) despite different setups and embodiments. The latter is the question that interested us in this thesis. It relates to the problem of optimizing the reproduction of the task under environmental constraints. The "How-to-imitate?" question is subdivided into two problems. The first problem, also known as the "correspondence problem", relates to resolving the discrepancy between the human demonstrator and robot's body that prevent the robot from doing an identical reproduction of the task. Even though we helped ourselves by considering solely humanoid platforms, that is platforms that have a joint configuration similar to that of the human, discrepancies in the number of degrees of freedom and range of motion remained. We resolved these by exploiting the redundant information conveyed through the demonstrations by collecting data through different frames of reference. By exploiting these redundancies in an algorithm comparable to the damped least square algorithm, we are able to reproduce a trajectory that minimizes the error between the desired trajectory and the reproduced trajectory across each frame of reference. The second problem consists in reproducing a trajectory in an unknown setup while respecting the task constraints learned during training. When the information learned from the demonstration no longer suffice to generalize the task constraints to a new set-up, the robot must re-learn the task; this time through trial-and-error. Here we considered the combination of trial-and-error learning to complement RbD. By adding a trial-and-error module to the original Imitation Learning algorithm, the robot can find a solution that is more adapted to the context and to its embodiment than the solution found using RbD. Specifically, we compared Reinforcement Learning (RL) – to other classical optimization techniques. We show that the system is advantageous in that: a) learning is more robust to unexpected events that have not been encountered during the demonstrations and b) the robot is able to optimize its own model of the task according to its own embodiment.

EPFL2009