Developing multi-agent adversarial environment using reinforcement learning and imitation learning

Robots had a great impact on the manufacturing industry ever since the early seventies when companies such as KUKA and ABB started deploying their first industrial robots. These robots merely performed very specific tasks in specific ways within well-defined environments. Still, they proved to be very useful as they could exceed human performance at these tasks. However, in order to enable robots to enter our daily life, they need to become more versatile and need to operate in much less structured environments. This thesis is partly devoted to stretching these limitations by means of learning, namely imitation learning (IL) and inverse reinforcement learning (IRL). Reinforcement learning (RL) is a powerful approach to enable robots to solve a task in an unknown environment. The practitioner describes a desired behavior by specifying a reward function and the robot autonomously interacts with the environment in order to find a control policy that generates high accumulated reward. However, RL is not suitable for teaching new tasks by non-experts because specifying appropriate reward functions can be difficult. Demonstrating the desired behavior is often easier for non-experts. Imitation learning can be used in order to enable the robot to reproduce the demonstrations. However, without explicitly inferring and modeling the intentions of the demonstrations, it can become difficult to solve the task for unseen situations. Inverse reinforcement learning (IRL) therefore aims to infer a reward function from the demonstrations, such that optimizing this reward function yields the desired behavior even for different situations. This thesis introduces a unifying approach to solve the inverse reinforcement learning problem in the same way as the reinforcement learning problem. This is achieved by framing both problems as information projection problems, i.e., we strive to minimize the relative entropy between a probabilistic model of the robot behavior and a given desired distribution. Furthermore, a trust region on the robot behavior is used to stabilize the optimization. For inverse reinforcement learning, the desired distribution is implicitly given by the expert demonstrations. The resulting optimization can be efficiently solved using state-of-the-art reinforcement learning methods. For reinforcement learning, the log-likelihood of the desired distribution is given by the reward function. The resulting optimization problem corresponds to a standard reinforcement learning formulation, except for an additional objective of maximizing the entropy of the robot behavior. This entropy objective adds little overhead to the optimization, but can lead to better exploration and more diversified policies. Trust-region I-projections are not only useful for training robots, but can also be applied to other machine learning problems. I-projections are typically used for variational inference, in order to approximate an intractable distribution by a simpler model. However, the resulting optimization problems are usually optimized based on stochastic gradient descent which often suffers from high variance in the gradient estimates. As trust-region I-projections where shown to be effective for reinforcement learning and inverse reinforcement learning, this thesis also explores their use for variational inference. More specifically, trust-region I-projections are investigated for the problem of approximating an intractable distribution by a Gaussian mixture model (GMM) with an adaptive number of components. GMMs are highly desirable for variational inference because they can yield arbitrary accurate approximations while inference from GMMs is still relatively cheap. In order to make learning the GMM feasible, we derive a lower bound that enables us to decompose the objective function. The optimization can then be performed by iteratively updating individual components using a technique from reinforcement learning. The resulting method is capable of learning approximations of significantly higher quality than existing variational inference methods. Due to the similarity of the underlying optimization problems, the insights gained from our variational inference method are also useful for IL and IRL. Namely, a similar lower bound can be applied also for the I-projection formulation of imitation learning. However, whereas for variational inference the lower bound serves to decompose the objective function, for imitation learning it allows us to provide a reward signal to the robot that does not depend on its behavior. Compared to reward functions that are relative to the current behavior of the robot---which are typical for popular adversarial methods---behavior-independent reward functions have the advantages that we can show convergence even for greedy optimization. Furthermore, behavior-independent reward functions solve the inverse reinforcement learning problem, thereby closing the gap between imitation learning and IRL. However, algorithms derived from our non-adversarial formulation are actually very similar to existing AIL methods, and we can even show that adversarial inverse reinforcement learning (AIRL) is indeed an instance of our formulation. AIRL was derived from an adversarial formulation, and we point out several problems of that derivation. In contrast, we show that AIRL can be straightforwardly derived from out non-adversarial formulation. Furthermore, we demonstrate that the non-adversarial formulation can be also used to derive novel algorithms by presenting a non-adversarial method for offline imitation learning.

This paper investigates the error accumulation problem in centralized training and decentralized execution (CTDE) policy‐based multi‐agent reinforcement learning (MARL) algorithms, which arises from local observation inaccuracies. To address this issue, we propose a novel MARL algorithm that incorporates imitation learning using local observations. Firstly, by analysing the multi‐agent proximal policy optimization algorithm and examining the problems arising when global states are replaced with local observations, it is proved that insufficient observations can lead to information loss, thereby introducing errors of advantage function, and it is demonstrated that the generalized advantage estimation method accumulates errors during the training process. Then, imitation learning is introduced and a novel training framework that combines reinforcement learning and imitation learning is proposed. During the reinforcement learning phase, an MARL agent trained with global observations acts as an expert. Subsequently, imitation learning is applied to train another agent that mimics the expert's decisions using only local observations. Finally, the effectiveness of this algorithm is verified in some commonly used multi‐agent environments, which demonstrates its superior performance compared to traditional multi‐agent reinforcement learning algorithms.

Chemical space is routinely explored by machine learning methods to discover interesting molecules, before time-consuming experimental synthesizing is attempted. However, these methods often rely on a graph representation, ignoring 3D information necessary for determining the stability of the molecules. We propose a reinforcement learning (RL) approach for generating molecules in Cartesian coordinates allowing for quantum chemical prediction of the stability. To improve sample-efficiency we learn basic chemical rules from imitation learning (IL) on the GDB-11 database to create an initial model applicable for all stoichiometries. We then deploy multiple copies of the model conditioned on a specific stoichiometry in a RL setting. The models correctly identify low energy molecules in the database and produce novel isomers not found in the training set. Finally, we apply the model to larger molecules to show how RL further refines the IL model in domains far from the training data.

The UAV pursuit-evasion strategy based on Deep Deterministic Policy Gradient (DDPG) algorithm is a current research hotspot. However, this algorithm has the defect of low efficiency in sample exploration. To solve this problem, this paper uses the imitation learning (IL) to improve the DDPG exploration strategy. A kind of quasiproportional guidance control law is designed to generate effective learning samples, which are used as the data of the initial experience pool of DDPG algorithm. The UAV pursuit-evasion strategy based on DDPG and imitation learning (IL-DDPG) is proposed, and the algorithm obtains the data from the experience pool for experience playback learning, which improves the exploration efficiency of the algorithm in the initial stage of training and avoids the problem of too many useless exploration in the training process. The simulation results show that the trained pursuit-UAV can flexibly adjust the flight speed and flight attitude to pursuit the evasion-UAV quickly. It also verifies that the improved DDPG algorithm is more effective than the basic DDPG algorithm to improve the training efficiency.

As robots and other intelligent agents move from simple environments and problems to more complex, unstructured settings, manually programming their behavior has become increasingly challenging and expensive. Often, it is easier for a teacher to demonstrate a desired behavior rather than attempt to manually engineer it. This process of learning from demonstrations, and the study of algorithms to do so, is called imitation learning. This work provides an introduction to imitation learning. It covers the underlying assumptions, approaches, and how they relate; the rich set of algorithms developed to tackle the problem; and advice on effective tools and implementation. We intend this paper to serve two audiences. First, we want to familiarize machine learning experts with the challenges of imitation learning, particularly those arising in robotics, and the interesting theoretical and practical distinctions between it and more familiar frameworks like statistical supervised learning theory and reinforcement learning. Second, we want to give roboticists and experts in applied artificial intelligence a broader appreciation for the frameworks and tools available for imitation learning. We organize our work by dividing imitation learning into directly replicating desired behavior (sometimes called behavioral cloning [Bain and Sammut, 1996]) and learning the hidden objectives of the desired behavior from demonstrations (called inverse optimal control [Kalman, 1964] or inverse reinforcement learning [Russell, 1998]). In addition to method analysis, we discuss the design decisions a practitioner must make when selecting an imitation learning approach. Moreover, application examples—such as robots that play table tennis Kober and Peters, 2009 and programs that play the game of Go Silver et al. 2016—illustrate the properties and motivations behind different forms of imitation learning. We conclude by presenting a set of open questions and point towards possible future research directions.

Model predictive control (MPC) has established itself as the standard advanced process control method. However, solving the resulting optimization problems in real-time can be challenging, especially when uncertainty is explicitly considered in a robust nonlinear predictive control approach. An increasingly popular alternative to avoid the online solution of the resulting optimization problems is to approximate their solution using neural networks. The networks are trained using many solutions of the MPC problem for different system states and therefore this approach is often called imitation learning. Controllers obtained via imitation learning have important drawbacks, since it is difficult to learn behaviors that are not well represented in the data and they must be redesigned from scratch when the control task changes. In this work, we show that these two drawbacks can be mitigated by combining imitation learning and concepts from reinforcement learning. The central idea is to use imitation learning as a very good initialization of a control policy that is iteratively updated using reinforcement learning, taking advantage of the fact that an explicit and differentiable expression of the approximate MPC controller is available. The efficacy of the combination of the two learning paradigms is highlighted via simulations of a semi-batch industrial polymerization reactor.

Generating paraphrases from given sentences involves decoding words step by step from a large vocabulary. To learn a decoder, supervised learning which maximizes the likelihood of tokens always suffers from the exposure bias. Although both reinforcement learning (RL) and imitation learning (IL) have been widely used to alleviate the bias, the lack of direct comparison leads to only a partial image on their benefits. In this work, we present an empirical study on how RL and IL can help boost the performance of generating paraphrases, with the pointer-generator as a base model. Experiments on the benchmark datasets show that (1) imitation learning is constantly better than reinforcement learning; and (2) the pointer-generator models with imitation learning outperform the state-of-the-art methods with a large margin.

Autonomous driving technology is constantly developing to a higher level of complex scenes, and there is a growing demand for the utilization of end-to-end data-driven control. However, the end-to-end path tracking process often encounters challenges in learning efficiency and generalization. To address this issue, this paper designs a deep deterministic policy gradient (DDPG)-based reinforcement learning strategy that integrates imitation learning and feedforward exploration in the path following process. In imitation learning, the path tracking control data generated by the model predictive control (MPC) method is used to train an end-to-end steering control model of a deep neural network. Another feedforward exploration behavior is predicted by road curvature and vehicle speed, and adds it and imitation learning to the DDPG reinforcement learning to obtain decision-making experience and action prediction behavior of the path tracking process. In the reinforcement learning process, imitation learning is used to update the pre-training parameters of the actor network, and a feedforward steering technique with random noise is adopted for strategy exploration. In the reward function, a hierarchical progressive reward form and a constrained objective reward function referring to MPC are designed, and the actor-critic network architecture is determined. Finally, the path tracking performance of the designed method is verified by comparing various training results, simulations, and HIL tests. The results show that the designed method can effectively utilize pre-training and feedforward prior experience to obtain optimal path tracking performance of an autonomous vehicle, and has better generalization ability than other methods. This study provides an efficient control scheme for improving the end-to-end control performance of autonomous vehicles.

Learning policies that generalize across multiple tasks is an important and challenging research topic in reinforcement learning and robotics. Training individual policies for every single potential task is often impractical, especially for continuous task variations, requiring more principled approaches to share and transfer knowledge among similar tasks. We present a novel approach for learning a nonlinear feedback policy that generalizes across multiple tasks. The key idea is to define a parametrized policy as a function of both the state and the task, which allows learning a single policy that generalizes across multiple known and unknown tasks. Applications of our novel approach to reinforcement and imitation learning in real-robot experiments are shown.

One of the major obstacles that hinders the application of robots to human day-to-day tasks is the current lack of flexible learning methods to endow the robots with the necessary skills and to allow them to adapt to new situations. In this work, we present a new intuitive method for teaching a robot anthropomorphic motion primitives. Our method combines the advantages of reinforcement and imitation learning in a single coherent framework. In contrast to existing approaches that use human demonstrations merely as an initialization for reinforcement learning, our method treats both ways of learning as homologous modules and chooses the most appropriate one in every situation. We apply Gaussian Process Regression to generalize a measure of value across the combined state-action-space. Based on the expected value, uncertainty, and expected deviation of generalized movements, our method decides whether to ask for a human demonstration or to improve its performance on its own, using reinforcement learning. The latter employs a probabilistic search strategy, based on expected deviation, that greatly accelerates learning while protecting the robot from unpredictable movements at the same time. To evaluate the performance of our approach, we conducted a series of experiments and successfully trained a robot to grasp an object at arbitrary positions on a table.

We propose a framework based on imitation learning and self-learning to enable robots to learn, improve, and generalize motor skills. The peg-in-hole task is important in manufacturing assembly work. Two motor skills for the peg-in-hole task are targeted: “hole search” and “peg insertion”. The robots learn initial motor skills from human demonstrations and then improve and/or generalize them through reinforcement learning (RL). An initial motor skill is represented as a concatenation of the parameters of a hidden Markov model (HMM) and a dynamic movement primitive (DMP) to classify input signals and generate motion trajectories. Reactions are classified as familiar or unfamiliar (i.e., modeled or not modeled), and initial motor skills are improved to solve familiar reactions and generalized to solve unfamiliar reactions. The proposed framework includes processes, algorithms, and reward functions that can be used for various motor skill types. To evaluate our framework, the motor skills were performed using an actual robotic arm and two reward functions for RL. To verify the learning and improving/generalizing processes, we successfully applied our framework to different shapes of pegs and holes. Moreover, the execution time steps and path optimization of RL were evaluated experimentally.

This paper proposes an intelligent game confrontation model for a wargame based on imitation learning and deep reinforcement learning, given light of the supremacy of reinforcement learning for training in complex environments. In the context of simulating the battle between the red team and the blue team, a pre-training model based on expert empirical data is produced using imitation learning, and a TD3 algorithm based on the attention mechanism is further designed to build an experience pool using priority experience replay. Finally, the model is enhanced using the self-play approach to increase its training efficiency. Experiments conducted after training demonstrate that the model has a superior training impact, and the winning rate in simulation training is enhanced by 8% compared to the original model.

Manycore chips are widely employed in high-performance computing and large-scale data analysis. However, the design of high-performance manycore chips is dominated by power and thermal constraints. In this respect, voltage-frequency island (VFI) is a promising design paradigm to create scalable energy-efficient platforms. By dynamically tailoring the voltage and frequency of each island, we can further improve the energy savings within given performance constraints. Inspired by the recent success of imitation learning (IL) in many application domains and its significant advantages over reinforcement learning (RL), we propose the first architecture-independent IL-based methodology for dynamic VFI (DVFI) control in manycore systems. Due to its popularity in the EDA community, we consider an RL-based DVFI control methodology as a strong baseline. Our experimental results demonstrate that IL is able to obtain higher quality policies than RL (on average, 5% less energy with the same level of performance) with significantly less computation time and hardware area overheads (3.1X and 8.8X, respectively).

Imitation learning (IL) is a popular paradigm for training policies in robotic systems when specifying the reward function is difficult. However, despite the success of IL algorithms, they impose the somewhat unrealistic requirement that the expert demonstrations must come from the same domain in which a new imitator policy is to be learned. We consider a practical setting, where (i) state-only expert demonstrations from the real (deployment) environment are given to the learner, (ii) the imitation learner policy is trained in a simulation (training) environment whose transition dynamics is slightly different from the real environment, and (iii) the learner does not have any access to the real environment during the training phase beyond the batch of demonstrations given. Most of the current IL methods, such as generative adversarial imitation learning and its state-only variants, fail to imitate the optimal expert behavior under the above setting. By leveraging insights from the Robust reinforcement learning (RL) literature and building on recent adversarial imitation approaches, we propose a robust IL algorithm to learn policies that can effectively transfer to the real environment without fine-tuning. Furthermore, we empirically demonstrate on continuous-control benchmarks that our method outperforms the state-of-the-art state-only IL method in terms of the zero-shot transfer performance in the real environment and robust performance under different testing conditions.

Reinforcement learning (RL) was first demonstrated to be a feasible approach to controlling heating, ventilation, and air conditioning (HVAC) systems more than a decade ago. However, there has been limited progress towards a practical and scalable RL solution for HVAC control. While one can train an RL agent in simulation, it is not cost-effective to create a model for each thermal zone or building. Likewise, existing RL agents generally take a long time to learn and are opaque to expert interrogation, making them unattractive for real-world deployment. To tackle these challenges, we propose Gnu-RL: a novel approach that enables practical deployment of RL for HVAC control and requires no prior information other than historical data from existing HVAC controllers. To achieve this, Gnu-RL adopts a recently-developed Differentiable Model Predictive Control (MPC) policy, which encodes domain knowledge on planning and system dynamics, making it both sample-efficient and interpretable. Prior to any interaction with the environment, a Gnu-RL agent is pre-trained on historical data using imitation learning, which enables it to match the behavior of the existing controller. Once it is put in charge of controlling the environment, the agent continues to improve its policy end-to-end, using a policy gradient algorithm. We evaluate Gnu-RL on both an EnergyPlus model and a real-world testbed. In both experiments, our agents were directly deployed in the environment after offline pre-training on expert demonstration. In the simulation experiment, our approach saved 6.6% energy compared to the best published RL result for the same environment, while maintaining a higher level of occupant comfort. Next, Gnu-RL was deployed to control the HVAC of a real-world conference room for a three-week period. Our results show that Gnu-RL saved 16.7% of cooling demand compared to the existing controller and tracked temperature set-point better.