Modern reinforcement learning systems produce many high-quality policies throughout the learning process. However, to choose which policy to actually deploy in the real world, they must be tested under an intractable number of environmental conditions. We introduce RPOSST, an algorithm to select a small set of test cases from a larger pool based on a relatively small number of sample evaluations. RPOSST treats the test case selection problem as a two-player game and optimizes a solution with provable $k$-of-$N$ robustness, bounding the error relative to a test that used all the test cases in the pool. Empirical results demonstrate that RPOSST finds a small set of test cases that identify high quality policies in a toy one-shot game, poker datasets, and a high-fidelity racing simulator.

Many prediction, decision-making, and control architectures rely on online learned Gaussian process (GP) models. However, most existing GP regression algorithms assume a single generative model, leading to poor predictive performance when the data are nonstationary, i.e., generated from multiple switching processes. Furthermore, existing methods for GP regression over nonstationary data require significant computation, do not come with provable guarantees on correctness and speed, and many only work in batch settings, making them ill-suited for real-time prediction. We present an efficient online GP framework, GP-non-Bayesian clustering (GP-NBC), which addresses these computational and theoretical issues, allowing for real-time changepoint detection and regression using GPs. Our empirical results on two real-world data sets and two synthetic data set show that GP-NBC outperforms state-of-the-art methods for nonstationary regression in terms of both regression error and computation. For example, it outperforms Dirichlet process GP clustering with Gibbs sampling by 98% in computation time reduction while the mean absolute error is comparable.

Reinforcement learning (RL) can be a tool for designing policies and controllers for robotic systems. However, the cost of real-world samples remains prohibitive as many RL algorithms require a large number of samples before learning useful policies. Simulators are one way to decrease the number of required real-world samples, but imperfect models make deciding when and how to trust samples from a simulator difficult. We present a framework for efficient RL in a scenario where multiple simulators of a target task are available, each with varying levels of fidelity. The framework is designed to limit the number of samples used in each successively higher-fidelity/cost simulator by allowing a learning agent to choose to run trajectories at the lowest level simulator that will still provide it with useful information. Theoretical proofs of the framework's sample complexity are given and empirical results are demonstrated on a remote-controlled car with multiple simulators. The approach enables RL algorithms to find near-optimal policies in a physical robot domain with fewer expensive real-world samples than previous transfer approaches or learning without simulators.

Learning from demonstration provides an attractive solution to the problem of teaching autonomous systems how to perform complex tasks. Reward learning from demonstration is a promising method of inferring a rich and transferable representation of the demonstrator's intents, but current algorithms suffer from intractability and inefficiency in large domains due to the assumption that the demonstrator is maximizing a single reward function throughout the whole task. This paper takes a different perspective by assuming that the reward function behind an unsegmented demonstration is actually composed of several distinct subtasks chained together. Leveraging this assumption, a Bayesian nonparametric reward-learning framework is presented that infers multiple subgoals and reward functions within a single unsegmented demonstration. The new framework is developed for discrete state spaces and also general continuous demonstration domains using Gaussian process reward representations. The algorithm is shown to have both performance and computational advantages over existing inverse reinforcement learning methods. Experimental results are given in both cases, demonstrating the ability to learn challenging maneuvers from demonstration on a quadrotor and a remote-controlled car.

This paper derives sample complexity results for using Gaussian Processes (GPs) in both model-based and model-free reinforcement learning (RL). We show that GPs are KWIK learnable, proving for the first time that a model-based RL approach using GPs, GP-Rmax, is sample efficient (PAC-MDP). However, we then show that previous approaches to model-free RL using GPs take an exponential number of steps to find an optimal policy, and are therefore not sample efficient. The third and main contribution is the introduction of a model-free RL algorithm using GPs, DGPQ, which is sample efficient and, in contrast to model-based algorithms, capable of acting in real time, as demonstrated on a five-dimensional aircraft simulator.

We present a framework for reinforcement learning (RL) in a scenario where multiple simulators are available with decreasing amounts of fidelity to the real-world learning scenario. Our framework is designed to limit the number of samples used in each successively higher-fidelity/cost simulator by allowing the agent to choose to run trajectories at the lowest level that will still provide it with information. The approach transfers state-action Q-values from lower-fidelity models as heuristics for the “Knows What It Knows” family of RL algorithms, which is applicable over a wide range of possible dynamics and reward representations. Theoretical proofs of the framework's sample complexity are given and empirical results are demonstrated on a remote controlled car with multiple simulators. The approach allows RL algorithms to find near-optimal policies for the real world with fewer expensive real-world samples than previous transfer approaches or learning without simulators.

This paper considers the problem of learning task specific web-service descriptions from traces of users successfully completing a task. Unlike prior approaches, we take a traditional machine-learning perspective to the construction of web-service mo

We propose an intelligent tutoring system that constructs a curriculum of hints and problems in order to teach a student skills with a rich dependency structure. We provide a template for building a multi-layered Dynamic Bayes Net to model this problem and describe how to learn the parameters of the model from data. Planning with the DBN then produces a teaching policy for the given domain. We test this end-to-end curriculum design system in two human-subject studies in the areas of finite field arithmetic and artificial language and show this method performs on par with hand-tuned expert policies.

We present theoretical and empirical results for a framework that combines the benefits of apprenticeship and autonomous reinforcement learning. Our approach modifies an existing apprenticeship learning framework that relies on teacher demonstrations and does not necessarily explore the environment. The first change is replacing previously used Mistake Bound model learners with a recently proposed framework that melds the KWIK and Mistake Bound supervised learning protocols. The second change is introducing a communication of expected utility from the student to the teacher. The resulting system only uses teacher traces when the agent needs to learn concepts it cannot efficiently learn on its own.

Currently, most systems for human-robot teaching allow only one mode of teacher-student interaction (e.g., teaching by demonstration or feedback), and teaching episodes have to be carefully set-up by an expert. To understand how we might integrate multiple, interleaved forms of human instruction into a robot learner, we performed a behavioral study in which 44 untrained humans were allowed to freely mix interaction modes to teach a simulated robot (secretly controlled by a human) a complex task. Analysis of transcripts showed that human teachers often give instructions that are nontrivial to interpret and not easily translated into a form useable by machine learning algorithms. In particular, humans often use implicit instructions, fail to clearly indicate the boundaries of procedures, and tightly interleave testing, feedback, and new instruction. In this paper, we detail these teaching patterns and discuss the challenges they pose to automatic teaching interpretation as well as the machine-learning algorithms that must ultimately process these instructions. We highlight the challenges by demonstrating the difficulties of an initial automatic teacher interpretation system.

Recent advancements in model-based reinforcement learning have shown that the dynamics of many structured domains (e.g. DBNs) can be learned with tractable sample complexity, despite their exponentially large state spaces. Unfortunately, these algorithms all require access to a planner that computes a near optimal policy, and while many traditional MDP algorithms make this guarantee, their computation time grows with the number of states. We show how to replace these over-matched planners with a class of sample-based planners — whose computation time is independent of the number of states — without sacrificing the sample-efficiency guarantees of the overall learning algorithms. To do so, we define sufficient criteria for a sample-based planner to be used in such a learning system and analyze two popular sample-based approaches from the literature. We also introduce our own sample-based planner, which combines the strategies from these algorithms and still meets the criteria for integration into our learning system. In doing so, we define the first complete RL solution for compactly represented (exponentially sized) state spaces with efficiently learnable dynamics that is both sample efficient and whose computation time does not grow rapidly with the number of states.

We introduce a learning framework that combines elements of the well-known PAC and mistake-bound models. The KWIK (knows what it knows) framework was designed particularly for its utility in learning settings where active exploration can impact the training examples the learner is exposed to, as is true in reinforcement-learning and active-learning problems. We catalog several KWIK-learnable classes and open problems.

Previous algorithms for learning lexicographic preference models (LPMs) produce a "best guess" LPM that is consistent with the observations. Our approach is more democratic: we do not commit to a single LPM. Instead, we approximate the target using the votes of a collection of consistent LPMs. We present two variations of this method---variable voting and model voting---and empirically show that these democratic algorithms outperform the existing methods. We also introduce an intuitive yet powerful learning bias to prune some of the possible LPMs. We demonstrate how this learning bias can be used with variable and model voting and show that the learning bias improves the learning curve significantly, especially when the number of observations is small.