Search

Approximate policy iteration is a class of reinforcement learning (RL) algorithms where the policy is encoded using a function approximator and which has been especially prominent in RL with continuous action spaces. In this class of RL algorithms, ensuring increase of the policy return during policy update often requires to constrain the change in action distribution. Several approximations exist in the literature to solve this constrained policy update problem. In this paper, we propose to improve over such solutions by introducing a set of projections that transform the constrained problem into an unconstrained one which is then solved by standard gradient descent. Using these projections, we empirically demonstrate that our approach can improve the policy update solution and the control over exploration of existing approximate policy iteration algorithms.

Many of the recent trajectory optimization algorithms alternate between linear approximation of the system dynamics around the mean trajectory and conservative policy update. One way of constraining the policy change is by bounding the Kullback-Leibler (KL) divergence between successive policies. These approaches already demonstrated great experimental success in challenging problems such as end-to-end control of physical systems. However, these approaches lack any improvement guarantee as the linear approximation of the system dynamics can introduce a bias in the policy update and prevent convergence to the optimal policy. In this article, we propose a new model-free trajectory-based policy optimization algorithm with guaranteed monotonic improvement. The algorithm backpropagates a local, quadratic and time-dependent Q-Function learned from trajectory data instead of a model of the system dynamics. Our policy update ensures exact KL-constraint satisfaction without simplifying assumptions on the system dynamics. We experimentally demonstrate on highly non-linear control tasks the improvement in performance of our algorithm in comparison to approaches linearizing the system dynamics. In order to show the monotonic improvement of our algorithm, we additionally conduct a theoretical analysis of our policy update scheme to derive a lower bound of the change in policy return between successive iterations.

The value function, at the core of the Bellmanian Reinforcement Learning framework, associates to each state the discounted expected cumulative reward which can be gathered after visiting this state. Given an (optimal) value function, an (optimal) policy is most simply derived by "greedification", heading in each time step toward the neighbor state with maximal value.Following the Bellman equations, the value function can be built by (approximate) dynamic programming, albeit facing severe scalability limitations in large state and action spaces. An alternative, inspired from the Energy-based learning framework (LeCun et al. 2006), is investigated in this paper, searching for a pseudo-value function such that it induces the same local order on the state space as a (nearly) optimal value function. By construction, the greedification of such a pseudo-value induces the same policy as the value function itself. The presented Direct Value Learning (DiVa) approach proceeds by directly learning the pseudo-value, taking some inspiration from the Inverse Reinforcement Learning (IRL) approach. In IRL, expert demonstrations are used to infer the reward function. Quite the contrary, DiVa uses bad demonstrations to infer the pseudo-value. Bad demonstrations are notoriously easier to generate than expert ones; typically, applying a random policy on a good initial state (e.g., a bicycle in equilibrium) will on average lead to visit states with decreasing values (the bicycle ultimately falls down). DiVa thus uses bad demonstrations, generated from weak prior knowledge, to learn a pseudo-value along a standard learning-to-rank approach. The derived pseudo-value directly induces a policy in the model-based RL framework, when the transition function is known. In the model-free RL setting, the state pseudo-value is exploited using off-policy learning, to infer a state-action pseudo-value and induce a policy.The proposed DiVa approach and the use of bad demonstrations to achieve direct value learning is original to our best knowledge. The loss of optimality of the pseudo value-based policy is analyzed and it is shown that it is bounded under mild assumptions. Finally, the experimental validation of DiVa on the mountain car, the bicycle and the swing-up pendulum problems comparatively demonstrates the simplicity and the merits of the approach.

The value function, at the core of the Bellmanian Reinforcement Learning framework, associates to each state the discounted expected cumulative reward which can be gathered after visiting this state. Given an (optimal) value function, an (optimal) policy is most simply derived by "greedification", heading in each time step toward the neighbor state with maximal value.Following the Bellman equations, the value function can be built by (approximate) dynamic programming, albeit facing severe scalability limitations in large state and action spaces. An alternative, inspired from the Energy-based learning framework (LeCun et al. 2006), is investigated in this paper, searching for a pseudo-value function such that it induces the same local order on the state space as a (nearly) optimal value function. By construction, the greedification of such a pseudo-value induces the same policy as the value function itself. The presented Direct Value Learning (DiVa) approach proceeds by directly learning the pseudo-value, taking some inspiration from the Inverse Reinforcement Learning (IRL) approach. In IRL, expert demonstrations are used to infer the reward function. Quite the contrary, DiVa uses bad demonstrations to infer the pseudo-value. Bad demonstrations are notoriously easier to generate than expert ones; typically, applying a random policy on a good initial state (e.g., a bicycle in equilibrium) will on average lead to visit states with decreasing values (the bicycle ultimately falls down). DiVa thus uses bad demonstrations, generated from weak prior knowledge, to learn a pseudo-value along a standard learning-to-rank approach. The derived pseudo-value directly induces a policy in the model-based RL framework, when the transition function is known. In the model-free RL setting, the state pseudo-value is exploited using off-policy learning, to infer a state-action pseudo-value and induce a policy.The proposed DiVa approach and the use of bad demonstrations to achieve direct value learning is original to our best knowledge. The loss of optimality of the pseudo value-based policy is analyzed and it is shown that it is bounded under mild assumptions. Finally, the experimental validation of DiVa on the mountain car, the bicycle and the swing-up pendulum problems comparatively demonstrates the simplicity and the merits of the approach.

International audience; Taking inspiration from inverse reinforcement learning, the proposed Direct Value Learning for Reinforcement Learning (DIVA) approach uses light priors to gener-ate inappropriate behaviors, and uses the corresponding state sequences to directly learn a value function. When the transition model is known, this value function directly defines a (nearly) optimal controller. Otherwise, the value function is extended to the state-action space using off-policy learning. The experimental validation of DIVA on the mountain car problem shows the robustness of the approach comparatively to SARSA, based on the assumption that the target state is known. The experimental validation on the bicycle problem shows that DIVA still finds good policies when relaxing this assumption.

The thesis contributions resolves around sequential decision taking and more precisely Reinforcement Learning (RL). Taking its root in Machine Learning in the same way as supervised and unsupervised learning, RL quickly grow in popularity within the last two decades due to a handful of achievements on both the theoretical and applicative front. RL supposes that the learning agent and its environment follow a stochastic Markovian decision process over a state and action space. The process is said of decision as the agent is asked to choose at each time step an action to take. It is said stochastic as the effect of selecting a given action in a given state does not systematically yield the same state but rather defines a distribution over the state space. It is said to be Markovian as this distribution only depends on the current state-action pair. Consequently to the choice of an action, the agent receives a reward. The RL goal is then to solve the underlying optimization problem of finding the behaviour that maximizes the sum of rewards all along the interaction of the agent with its environment. From an applicative point of view, a large spectrum of problems can be cast onto an RL one, from Backgammon (TD-Gammon, was one of Machine Learning first success giving rise to a world class player of advanced level) to decision problems in the industrial and medical world. However, the optimization problem solved by RL depends on the prevous definition of a reward function that requires a certain level of domain expertise and also knowledge of the internal quirks of RL algorithms. As such, the first contribution of the thesis was to propose a learning framework that lightens the requirements made to the user. The latter does not need anymore to know the exact solution of the problem but to only be able to choose between two behaviours exhibited by the agent, the one that matches more closely the solution. Learning is interactive between the agent and the user and resolves around the three main following points: i) The agent demonstrates a behaviour ii) The user compares it w.r.t. to the current best one iii) The agent uses this feedback to update its preference model of the user and uses it to find the next behaviour to demonstrate. To reduce the number of required interactions before finding the optimal behaviour, the second contribution of the thesis was to define a theoretically sound criterion making the trade-off between the sometimes contradicting desires of complying with the user's preferences and demonstrating sufficiently different behaviours. The last contribution was to ensure the robustness of the algorithm w.r.t. the feedback errors that the user might make. Which happens more often than not in practice, especially at the initial phase of the interaction, when all the behaviours are far from the expected solution.; Les contributions de la thèse sont centrées sur la prise de décisions séquentielles et plus spécialement sur l'Apprentissage par Renforcement (AR). Prenant sa source de l'apprentissage statistique au même titre que l'apprentissage supervisé et non-supervisé, l'AR a gagné en popularité ces deux dernières décennies en raisons de percées aussi bien applicatives que théoriques. L'AR suppose que l'agent (apprenant) ainsi que son environnement suivent un processus de décision stochastique Markovien sur un espace d'états et d'actions. Le processus est dit de décision parce que l'agent est appelé à choisir à chaque pas de temps du processus l'action à prendre. Il est dit stochastique parce que le choix d'une action donnée en un état donné n'implique pas le passage systématique à un état particulier mais définit plutôt une distribution sur l'espace d'états. Il est dit Markovien parce que cette distribution ne dépend que de l'état et de l'action courante. En conséquence d'un choix d'action, l'agent reçoit une récompense. Le but de l'AR est alors de résoudre le problème d'optimisation retournant le comportement qui assure à l'agent une récompense maximale tout au long de son interaction avec l'environnement. D'un point de vue pratique, un large éventail de problèmes peuvent être transformés en un problème d'AR, du Backgammon (cf. TD-Gammon, l'une des premières grandes réussites de l'AR et de l'apprentissage statistique en général, donnant lieu à un joueur expert de classe internationale) à des problèmes de décision dans le monde industriel ou médical. Seulement, le problème d'optimisation résolu par l'AR dépend de la définition préalable d'une fonction de récompense adéquate nécessitant une expertise certaine du domaine d'intérêt mais aussi du fonctionnement interne des algorithmes d'AR. En ce sens, la première contribution de la thèse a été de proposer un nouveau cadre d'apprentissage, allégeant les prérequis exigés à l'utilisateur. Ainsi, ce dernier n'a plus besoin de connaître la solution exacte du problème mais seulement de pouvoir désigner entre deux comportements, celui qui s'approche le plus de la solution. L'apprentissage se déroule en interaction entre l'utilisateur et l'agent. Cette interaction s'articule autour des trois points suivants : i) L'agent exhibe un nouveau comportement ii) l'expert le compare au meilleur comportement jusqu'à présent iii) l'agent utilise ce retour pour mettre à jour son modèle des préférences puis choisit le prochain comportement à démontrer. Afin de réduire le nombre d'interactions nécessaires entre l'utilisateur et l'agent pour que ce dernier trouve le comportement optimal, la seconde contribution de la thèse a été de définir un critère théoriquement justifié faisant le compromis entre les désirs parfois contradictoires de prendre en compte les préférences de l'utilisateur tout en exhibant des comportements suffisamment différents de ceux déjà proposés. La dernière contribution de la thèse est d'assurer la robustesse de l'algorithme face aux éventuelles erreurs d'appréciation de l'utilisateur. Ce qui arrive souvent en pratique, spécialement au début de l'interaction, quand tous les comportements proposés par l'agent sont loin de la solution attendue.

The thesis contributions resolves around sequential decision taking and more precisely Reinforcement Learning (RL). Taking its root in Machine Learning in the same way as supervised and unsupervised learning, RL quickly grow in popularity within the last two decades due to a handful of achievements on both the theoretical and applicative front. RL supposes that the learning agent and its environment follow a stochastic Markovian decision process over a state and action space. The process is said of decision as the agent is asked to choose at each time step an action to take. It is said stochastic as the effect of selecting a given action in a given state does not systematically yield the same state but rather defines a distribution over the state space. It is said to be Markovian as this distribution only depends on the current state-action pair. Consequently to the choice of an action, the agent receives a reward. The RL goal is then to solve the underlying optimization problem of finding the behaviour that maximizes the sum of rewards all along the interaction of the agent with its environment. From an applicative point of view, a large spectrum of problems can be cast onto an RL one, from Backgammon (TD-Gammon, was one of Machine Learning first success giving rise to a world class player of advanced level) to decision problems in the industrial and medical world. However, the optimization problem solved by RL depends on the prevous definition of a reward function that requires a certain level of domain expertise and also knowledge of the internal quirks of RL algorithms. As such, the first contribution of the thesis was to propose a learning framework that lightens the requirements made to the user. The latter does not need anymore to know the exact solution of the problem but to only be able to choose between two behaviours exhibited by the agent, the one that matches more closely the solution. Learning is interactive between the agent and the user and resolves around the three main following points: i) The agent demonstrates a behaviour ii) The user compares it w.r.t. to the current best one iii) The agent uses this feedback to update its preference model of the user and uses it to find the next behaviour to demonstrate. To reduce the number of required interactions before finding the optimal behaviour, the second contribution of the thesis was to define a theoretically sound criterion making the trade-off between the sometimes contradicting desires of complying with the user's preferences and demonstrating sufficiently different behaviours. The last contribution was to ensure the robustness of the algorithm w.r.t. the feedback errors that the user might make. Which happens more often than not in practice, especially at the initial phase of the interaction, when all the behaviours are far from the expected solution., Les contributions de la thèse sont centrées sur la prise de décisions séquentielles et plus spécialement sur l'Apprentissage par Renforcement (AR). Prenant sa source de l'apprentissage statistique au même titre que l'apprentissage supervisé et non-supervisé, l'AR a gagné en popularité ces deux dernières décennies en raisons de percées aussi bien applicatives que théoriques. L'AR suppose que l'agent (apprenant) ainsi que son environnement suivent un processus de décision stochastique Markovien sur un espace d'états et d'actions. Le processus est dit de décision parce que l'agent est appelé à choisir à chaque pas de temps du processus l'action à prendre. Il est dit stochastique parce que le choix d'une action donnée en un état donné n'implique pas le passage systématique à un état particulier mais définit plutôt une distribution sur l'espace d'états. Il est dit Markovien parce que cette distribution ne dépend que de l'état et de l'action courante. En conséquence d'un choix d'action, l'agent reçoit une récompense. Le but de l'AR est alors de résoudre le problème d'optimisation retournant le comportement qui assure à l'agent une récompense maximale tout au long de son interaction avec l'environnement. D'un point de vue pratique, un large éventail de problèmes peuvent être transformés en un problème d'AR, du Backgammon (cf. TD-Gammon, l'une des premières grandes réussites de l'AR et de l'apprentissage statistique en général, donnant lieu à un joueur expert de classe internationale) à des problèmes de décision dans le monde industriel ou médical. Seulement, le problème d'optimisation résolu par l'AR dépend de la définition préalable d'une fonction de récompense adéquate nécessitant une expertise certaine du domaine d'intérêt mais aussi du fonctionnement interne des algorithmes d'AR. En ce sens, la première contribution de la thèse a été de proposer un nouveau cadre d'apprentissage, allégeant les prérequis exigés à l'utilisateur. Ainsi, ce dernier n'a plus besoin de connaître la solution exacte du problème mais seulement de pouvoir désigner entre deux comportements, celui qui s'approche le plus de la solution. L'apprentissage se déroule en interaction entre l'utilisateur et l'agent. Cette interaction s'articule autour des trois points suivants : i) L'agent exhibe un nouveau comportement ii) l'expert le compare au meilleur comportement jusqu'à présent iii) l'agent utilise ce retour pour mettre à jour son modèle des préférences puis choisit le prochain comportement à démontrer. Afin de réduire le nombre d'interactions nécessaires entre l'utilisateur et l'agent pour que ce dernier trouve le comportement optimal, la seconde contribution de la thèse a été de définir un critère théoriquement justifié faisant le compromis entre les désirs parfois contradictoires de prendre en compte les préférences de l'utilisateur tout en exhibant des comportements suffisamment différents de ceux déjà proposés. La dernière contribution de la thèse est d'assurer la robustesse de l'algorithme face aux éventuelles erreurs d'appréciation de l'utilisateur. Ce qui arrive souvent en pratique, spécialement au début de l'interaction, quand tous les comportements proposés par l'agent sont loin de la solution attendue.

International audience; Aimed at on-board robot training, an approach hybridizing active preference learning and reinforcement learning is presented: Interactive Bayesian Policy Search (IBPS) builds a robotic controller through direct and frugal interaction with the human expert, iteratively emitting preferences among a few behaviors demonstrated by the robot. These preferences allow the robot to gradually refine its policy utility estimate, and select a new policy to be demonstrated, after an Expected Utility of Selection criterion. The paper contribution is on handling the preference noise, due to expert's mistakes or disinterest when demonstrated behaviors are equally unsatisfactory. A noise model is proposed, enabling a resource-limited robot to soundly estimate the preference noise and maintain a robust interaction with the expert, thus enforcing a low sample complexity. A proof of principle of the IBPS approach, in simulation and on-board, is presented.

International audience; Learning by imitation, among the most promising techniques for reinforcement learning in complex domains, critically depends on the human designer ability to provide sufficiently many demonstrations of satisfactory quality. The approach presented in this paper, referred to as DIVA (Direct Value Learning for Reinforcement Learning), aims at addressing both above limitations by exploiting simple experiments. The approach stems from a straightforward remark: while it is rather easy to set a robot in a target situation, the quality of its situation will naturally deteriorate upon the action of naive controllers. The demonstration of such naive controllers can thus be used to learn directly a value function, through a preference learning approach. Under some conditions on the transition model, this value function enables to define an optimal controller. The DIVA approach is experimentally demonstrated by teaching a robot to follow another robot. Importantly, the approach does not require any robotic simulator to be available, nor does it require any pattern-recognition primitive (e.g. seeing the other robot) to be provided.