Prioritized Sweeping Converges to the Optimal Value Function

Deep reinforcement learning is the technology of artificial neural networks in the field of decision-making and control. The traditional model-free reinforcement learning algorithm requires a large amount of environment interactive data to iterate the algorithm. This model’s performance also suffers due to low utilization of training data, while the model-based reinforcement learning (MBRL) algorithm improves the efficiency of the data, MBRL locks into low prediction accuracy. Although MBRL can utilize the additional data generated by the dynamic model, a system dynamics model with low prediction accuracy will provide low-quality data and affect the algorithm’s final result. In this paper, based on the A3C (Asynchronous Advantage Actor-Critic) algorithm, an improved model-based deep reinforcement learning algorithm using a learning degree network (MBRL-LDN) is presented. By comparing the differences between the predicted states outputted by the proposed multidynamic model and the original predicted states, the learning degree of the system dynamics model is calculated. The learning degree represents the quality of the data generated by the dynamic model and is used to decide whether to continue to interact with the dynamic model during a particular episode. Thus, low-quality data will be discarded. The superiority of the proposed method is verified by conducting extensive contrast experiments.

Although model-free reinforcement learning methods are illustrated to own abilities to learn a wide range of control and decision-making tasks, it is time-consuming and needs a huge number of samples to learn optimal policy. Therefore, model-based reinforcement learning algorithms are always treated as the most efficient ways to settle robotic control problems. In this work, a model-based reinforcement learning tracking control method with free energy is proposed. The control objective is to make sure that the proposed controller can let a robotic manipulator with parameters uncertainties and disturbance follow the desired trajectory and obtain good tracking performance. The main contribution of this work is to employ the free energy principle to optimize the problem of robotic control. The probability distribution of robotic dynamic model is learnt by a evolving neural network, a free energy based reinforcement learning control method uses the learnt model to find optimal control policy. The experiment results demonstrate that the proposed tracking control method is effective in solving robotic tracking control problems.

Reinforcement learning (RL) focuses on an essential aspect of intelligent behavior – how an agent can learn to make good decisions given experience and rewards in a stochastic<br>world. Yet popular RL algorithms that have enabled exciting successes in domains with good simulators (Go, Atari, etc) still often fail to learn in other domains because they rely on<br>simple heuristics for exploration. This provides additional empirical justification for essential questions around RL, specifically around algorithms that learn in a provably efficient manner through strategic exploration in any considered domain. This thesis provides new algorithms<br>and theory that enable good performance with respect to existing theoretical frameworks for evaluating RL algorithms (specifically, probably approximately correct) and introduces new stronger evaluation criteria, that may be particularly of interest as RL is applied to more real world problems.<br>For the first line of work on probably approximately correct (PAC) RL algorithms, we introduce a series of algorithms for episodic tabular domains with substantially better PAC<br>sample complexity bounds that culminate in a new algorithm with close to minimax optimal PAC and regret bounds. Look up tables are required by most sample efficient and computationally tractable algorithms, but cannot represent many practical domains. We therefore also present a new RL algorithm that can learn a good policy in environments with high dimensional observations and hidden deterministic states; unlike predecessors, this algorithm provably<br>explores not only in a statistically but also computationally efficient manner assuming access to function classes with efficient optimization oracles. To make progress it is critical to have the right measures of success. While empirical<br>demonstrations are quite clear, we find that for theoretical properties, two of the most commonly used learning frameworks, PAC guarantees and regret guarantees, each allow undesirable algorithm behavior (e.g. ignoring new observations that could improve the policy). We present<br>a new stronger learning framework called Uniform-PAC that unifies the existing frameworks and prevents undesirable algorithm properties. One caveat of all existing learning frameworks is that for any particular episode, we do not<br>know how well the algorithm will perform. To address this, we introduce the IPOC framework that requires algorithms to provide a certificate before each episode bounding how suboptimal the current policy can be. Such certifications may be of substantial interest in high stakes scenarios when an organization may wish to track or even pause an online RL system should the potential expected performance bound drop below a required expected outcome.

Arbitration between model-based (MB) and model-free (MF) reinforcement learning (RL) is key feature of human reinforcement learning. The computational model of arbitration control has been demonstrated to outperform conventional reinforcement learning algorithm, in terms of not only behavioral data but also neural signals. However, this arbitration process does not take full account of contextual changes in environment during learning. By incorporating a Dirichlet process Gaussian mixture model into the arbitration process, we propose a meta-controller for RL that quickly adapts to contextual changes of environment. The proposed model performs better than a conventional model-free RL, model-based RL, and arbitration model.

Over the last couple of decades the demand for high precision and enhanced performance of physical systems has been steadily increasing. This demand often results in miniaturization and complex design, thus increasing the need for complex nonlinear control methods. Some of the state of the art nonlinear methods are stymied by the requirement of full state information, model and parameter uncertainties, mathematical complexity, etc. For many scenarios it is nearly impossible to consider all the uncertainties during the design of a feedback controller. Additionally, while designing a modelbased nonlinear control there is no standard mechanism to incorporate performance measures. Some of the mentioned issues can be addressed by using online learning. Animals and humans have the ability to share, explore, act or respond, memorize the outcome and repeat the task to achieve a better outcome when they encounter the same or a similar scenario. This is called learning from interaction. One instance of this approach is reinforcement learning (RL). However, RL methods are hindered by the curse of dimensionality, non-interpretability and non-monotonic convergence of the learning algorithms. This can be attributed to the intrinsic characteristics of RL, as it is a modelfree approach and hence no standard mechanism exists to incorporate a priori model information. In this thesis, learning methods are proposed which explicitly use the available system knowledge. This can be seen as a new class of approaches that bridge model-based and model-free methods. These methods can address some of the hurdles mentioned earlier. For example, i) a prior system information can speed up the learning, ii) new control objectives can be achieved which otherwise would be extremely difficult to attain using only model-based methods, iii) physical meaning can be attributed to the learned controller. The developed approach is as follows: themodel of the given physical system is represented in the port-Hamiltonian (PH) form. For the system dynamics in PH form a passivity-based control (PBC) law is formulated, which often requires the solution to a set of partial differential equations (PDEs). Instead of finding an analytical solution, the PBC control law is parameterized using an unknown parameter vector. Then, by using a variation of the standard actor-critic learning algorithm, the unknown parameters can be learned online. Using the principles of stochastic approximation theory, a proof of convergence for the developed method is shown. The proposedmethods are evaluated for the stabilization and regulation ofmechanical and electro-mechanical systems. The simulation and experimental results show comparable learning curves. In the final part of the thesis a novel integral reinforcement learning approach is developed to solve for the optimal output tracking control problem for a set of linear heterogeneous multi-agent systems. Unlike existing methods, this approach does not need to solve either the output regulator equation or requires a p-copy of the leader’s dynamics in the agent’s control law. A detailed numerical evaluation has been conducted to show the feasibility of the developed method.

The Wisconsin Card Sorting Test (WCST) is considered a gold standard for the assessment of cognitive flexibility. On the WCST, repeating a sorting category following negative feedback is typically treated as indicating reduced cognitive flexibility. Therefore such responses are referred to as ‘perseveration’ errors. Recent research suggests that the propensity for perseveration errors is modulated by response demands: They occur less frequently when their commitment repeats the previously executed response. Here, we propose parallel reinforcement-learning models of card sorting performance, which assume that card sorting performance can be conceptualized as resulting from model-free reinforcement learning at the level of responses that occurs in parallel with model-based reinforcement learning at the categorical level. We compared parallel reinforcement-learning models with purely model-based reinforcement learning, and with the state-of-the-art attentional-updating model. We analyzed data from 375 participants who completed a computerized WCST. Parallel reinforcement-learning models showed best predictive accuracies for the majority of participants. Only parallel reinforcement-learning models accounted for the modulation of perseveration propensity by response demands. In conclusion, parallel reinforcement-learning models provide a new theoretical perspective on card sorting and it offers a suitable framework for discerning individual differences in latent processes that subserve behavioral flexibility.

To behave adaptively, we must learn from the consequences of our actions. Doing so is difficult when the consequences of an action follow a delay. This introduces the problem of temporal credit assignment. When feedback follows a sequence of decisions, how should the individual assign credit to the intermediate actions that comprise the sequence? Research in reinforcement learning provides 2 general solutions to this problem: model-free reinforcement learning and model-based reinforcement learning. In this review, we examine connections between stimulus-response and cognitive learning theories, habitual and goal-directed control, and model-free and model-based reinforcement learning. We then consider a range of problems related to temporal credit assignment. These include second-order conditioning and secondary reinforcers, latent learning and detour behavior, partially observable Markov decision processes, actions with distributed outcomes, and hierarchical learning. We ask whether humans and animals, when faced with these problems, behave in a manner consistent with reinforcement learning techniques. Throughout, we seek to identify neural substrates of model-free and model-based reinforcement learning. The former class of techniques is understood in terms of the neurotransmitter dopamine and its effects in the basal ganglia. The latter is understood in terms of a distributed network of regions including the prefrontal cortex, medial temporal lobes, cerebellum, and basal ganglia. Not only do reinforcement learning techniques have a natural interpretation in terms of human and animal behavior but they also provide a useful framework for understanding neural reward valuation and action selection.

Energy efficient speed planning of electric vehicles for car-following scenario using model-based reinforcement learning

Reinforcement learning (RL) is a simulation-based technique to solve Markov decision problems or processes (MDPs). It is especially useful if the transition probabilities in the MDP are hard to find or if the number of states in the problem is too large. In this paper, we present a new model-based RL algorithm that builds the transition probability model without the generation of the transition probabilities; the literature on model-based RL attempts to compute the transition probabilities. We also present a variance-penalized Bellman equation and an RL algorithm that uses it to solve a variance-penalized MDP. We conclude with some numerical experiments with these algorithms.

In Reinforcement Learning, an intelligent agent has to make a sequence of decisions to accomplish a goal. If this sequence is long, then the agent has to plan over a long horizon. While learning the optimal policy and its value function is a well studied problem in Reinforcement Learning, this paper focuses on the structure of the optimal value function and how hard it is to represent the optimal value function. We show that the generalized Rademacher complexity of the hypothesis space of all optimal value functions is dependent on the planning horizon and independent of the state and action space size. Further, we present bounds on the action-gaps of action value functions and show that they can collapse if a long planning horizon is used. The theoretical results are verified empirically on randomly generated MDPs and on a grid-world fruit collection task using deep value function approximation. Our theoretical results highlight a connection between value function approximation and the Options framework and suggest that value functions should be decomposed along bottlenecks of the MDP's transition dynamics.

Reinforcement learning (RL) for robotics is challenging due to the difficulty in hand-engineering a dense cost function, which can lead to unintended behavior, and dynamical uncertainty, which makes exploration and constraint satisfaction challenging. We address these issues with a new model-based reinforcement learning algorithm, Safety Augmented Value Estimation from Demonstrations (SAVED), which uses supervision that only identifies task completion and a modest set of suboptimal demonstrations to constrain exploration and learn efficiently while handling complex constraints. We then compare SAVED with 3 state-of-the-art model-based and model-free RL algorithms on 6 standard simulation benchmarks involving navigation and manipulation and a physical knot-tying task on the da Vinci surgical robot. Results suggest that SAVED outperforms prior methods in terms of success rate, constraint satisfaction, and sample efficiency, making it feasible to safely learn a control policy directly on a real robot in less than an hour. For tasks on the robot, baselines succeed less than 5% of the time while SAVED has a success rate of over 75% in the first 50 training iterations. Code and supplementary material is available at https://tinyurl.com/saved-rl.

Bidirectional Human-Robot Learning: Imitation and Skill Improvement

In this dissertation we develop numerical algorithms and software tools to facilitate parallel solutions of nonlinear programming (NLP) problems. In particular, we address large-scale, block-structured problems with an intrinsic decomposable configuration. These problems arise in a great number of engineering applications, including parameter estimation, optimal control, network optimization, and stochastic programming. The structure of these problems can be leveraged by optimization solvers to accelerate solutions and overcome memory limitations, and we propose variants to two classes of optimization algorithms: augmented Lagrangian (AL) schemes and Schur-complement interior-point methods. The convergence properties of augmented Lagrangian decomposition schemes like the alternating direction method of multipliers (ADMM) and progressive hedging (PH) are well established for convex optimization but convergence guarantees in non-convex settings are still poorly understood. In practice, however, ADMM and PH often perform satisfactorily in complex non-convex NLPs. In this work, we study connections between the method of multipliers (MM), ADMM, and PH to derive benchmarking metrics that explain why PH and ADMM work in practice. We illustrate the concepts using challenging dynamic optimization problems. Our exposition seeks to establish more formalism in benchmarking ADMM, PH, and AL schemes and to motivate algorithmic improvements. The effectiveness of nonlinear interior-point solvers for solving large-scale problems relies quite heavily on the solution of the underlying linear algebra systems. The schur-complement decomposition is very effective for parallelizing the solution of linear systems with modest coupling. However, for systems with large number of coupling variables the schur-complement method does not scale favorably. We implement an approach that uses a Krylov solver (GMRES) preconditioned with ADMM to solve block-structured linear systems that arise in the interior-point method. We show that this ADMM-GMRES approach overcomes the well-known scalability issues of Schur decomposition. One important drawback of using decomposition approaches like ADMM and PH is their convergence rate. Unlike Schur-complement interior-point algorithms that have super-linear convergence, augmented Lagrangian approaches typically exhibit linear and sublinear rates. We exploit connections between ADMM and the Schur-complement decomposition to derive an accelerated version of ADMM. Specifically, we study the effectiveness of performing a Newton-Raphson algorithm to compute multiplier estimates for augmented Lagrangian methods. We demonstrate using two-stage stochastic programming problems that our multiplier update achieves convergence in fewer iterations for MM on general nonlinear problems. In the case of ADMM, the newton update significantly reduces the number of subproblem solves for convex quadratic programs (QPs). Moreover, we show that using newton multiplier updates makes the method robust to the selection of the penalty parameter. Traditionally, state-of-the-art optimization solvers are implemented in low-level programming languages. In our experience, the development of decomposition algorithms in these frameworks is challenging. They present a steep learning curve and can slow the development and testing of new numerical algorithms. To mitigate these challenges, we developed PyNumero, a new open source framework implemented in Python and C++. The package seeks to facilitate development of optimization algorithms for large-scale optimization within a high-level programming environment while at the same time minimizing the computational burden of using Python. The efficiency of PyNumero is illustrated by implementing algorithms for problems arising in stochastic programming and optimal control. Timing results are presented for both serial and parallel implementations. Our computational studies demonstrate that with the appropriate balance between compiled code and Python, efficient implementations of optimization algorithms are achievable in these high-level languages.

In complex real-world environments, traditional (tabular) techniques for solving Reinforcement Learning (RL) do not scale. Function approximation is needed, but unfortunately, existing approaches generally have poor convergence and optimality guarantees. Additionally, for the case of human environments, it is valuable to be able to leverage human input. In this paper we introduce Expanding Value Function Approximation (EVFA), a function approximation algorithm that returns the optimal value function given sufficient rounds. To leverage human input, we introduce a new human-agent interaction scheme, training regimens, which allow humans to interact with and improve agent learning in the setting of a machine learning game. In experiments, we show EVFA compares favorably to standard value approximation approaches. We also show that training regimens enable humans to further improve EVFA performance. In our user study, we find that non-experts are able to provide effective regimens and that they found the game fun.

Recent evidence in neuroscience and psychology suggests that a single reinforcement learning (RL) algorithm only accounts for less than 60% of the variance of human choice behavior in an uncertain and dynamic environment, where the amount of uncertainty in state-action-state transitions drift over time. The prediction performance further decreases when the size of the state space increases. We proposed a hierarchical context-dependent RL control framework that dynamically exerted control weights on model-based (MB) and multiple model-free (MF) RL strategies associated with different task goals. To properly assess the validity of the proposed method, we considered a two-stage Markov decision task (MDT) in which the three different types of context changed over time. We trained 57 different RL control models on a Caltech MDT data set; then, we assessed their prediction performance using a Bayesian model comparison. This large-scale computer simulation analysis revealed that the model providing the most accurate prediction was the version that implemented the competition between the MB and multiple goal-dependent MF RL strategies. The present study demonstrates the applicability of the goal-driven RL control to a variety of real-world human-robot interaction scenarios.