Trust region methods rigorously enabled reinforcement learning (RL) agents to learn monotonically improving policies, leading to superior performance on a variety of tasks. Unfortunately, when it comes to multi-agent reinforcement learning (MARL), the property of monotonic improvement may not simply apply; this is because agents, even in cooperative games, could have conflicting directions of policy updates. As a result, achieving a guaranteed improvement on the joint policy where each agent acts individually remains an open challenge. In this paper, we extend the theory of trust region learning to MARL. Central to our findings are the multi-agent advantage decomposition lemma and the sequential policy update scheme. Based on these, we develop Heterogeneous-Agent Trust Region Policy Optimisation (HATPRO) and Heterogeneous-Agent Proximal Policy Optimisation (HAPPO) algorithms. Unlike many existing MARL algorithms, HATRPO/HAPPO do not need agents to share parameters, nor do they need any restrictive assumptions on decomposibility of the joint value function. Most importantly, we justify in theory the monotonic improvement property of HATRPO/HAPPO. We evaluate the proposed methods on a series of Multi-Agent MuJoCo and StarCraftII tasks. Results show that HATRPO and HAPPO significantly outperform strong baselines such as IPPO, MAPPO and MADDPG on all tested tasks, therefore establishing a new state of the art.

Reward functions are at the heart of every reinforcement learning (RL) algorithm. In robotic grasping, rewards are often complex and manually engineered functions that do not rely on well-justified physical models from grasp analysis. This work demonstrates that analytic grasp stability metrics constitute powerful optimization objectives for RL algorithms that refine grasps on a three-fingered hand using only tactile and joint position information. We outperform a binary-reward baseline by 42.9% and find that a combination of geometric and force-agnostic grasp stability metrics yields the highest average success rates of 95.4% for cuboids, 93.1% for cylinders, and 62.3% for spheres across wrist position errors between 0 and 7 centimeters and rotational errors between 0 and 14 degrees. In a second experiment, we show that grasp refinement algorithms trained with contact feedback (contact positions, normals, and forces) perform up to 6.6% better than a baseline that receives no tactile information.

In value-based deep reinforcement learning methods, approximation of value functions induces overestimation bias and leads to suboptimal policies. We show that in deep actor-critic methods that aim to overcome the overestimation bias, if the reinforcement signals received by the agent have a high variance, a significant underestimation bias arises. To minimize the underestimation, we introduce a parameter-free, novel deep Q-learning variant. Our Q-value update rule combines the notions behind Clipped Double Q-learning and Maxmin Q-learning by computing the critic objective through the nested combination of maximum and minimum operators to bound the approximate value estimates. We evaluate our modification on the suite of several OpenAI Gym continuous control tasks, improving the state-of-the-art in every environment tested.

Many real-life scenarios require humans to make difficult trade-offs: do we always follow all the traffic rules or do we violate the speed limit in an emergency? These scenarios force us to evaluate the trade-off between collective norms and our own personal objectives. To create effective AI-human teams, we must equip AI agents with a model of how humans make trade-offs in complex, constrained environments. These agents will be able to mirror human behavior or to draw human attention to situations where decision making could be improved. To this end, we propose a novel inverse reinforcement learning (IRL) method for learning implicit hard and soft constraints from demonstrations, enabling agents to quickly adapt to new settings. In addition, learning soft constraints over states, actions, and state features allows agents to transfer this knowledge to new domains that share similar aspects. We then use the constraint learning method to implement a novel system architecture that leverages a cognitive model of human decision making, multi-alternative decision field theory (MDFT), to orchestrate competing objectives. We evaluate the resulting agent on trajectory length, number of violated constraints, and total reward, demonstrating that our agent architecture is both general and achieves strong performance. Thus we are able to capture and replicate human-like trade-offs from demonstrations in environments when constraints are not explicit.

Reinforcement learning has driven impressive advances in machine learning. Simultaneously, quantum-enhanced machine learning algorithms using quantum annealing underlie heavy developments. Recently, a multi-agent reinforcement learning (MARL) architecture combining both paradigms has been proposed. This novel algorithm, which utilizes Quantum Boltzmann Machines (QBMs) for Q-value approximation has outperformed regular deep reinforcement learning in terms of time-steps needed to converge. However, this algorithm was restricted to single-agent and small 2x2 multi-agent grid domains. In this work, we propose an extension to the original concept in order to solve more challenging problems. Similar to classic DQNs, we add an experience replay buffer and use different networks for approximating the target and policy values. The experimental results show that learning becomes more stable and enables agents to find optimal policies in grid-domains with higher complexity. Additionally, we assess how parameter sharing influences the agents behavior in multi-agent domains. Quantum sampling proves to be a promising method for reinforcement learning tasks, but is currently limited by the QPU size and therefore by the size of the input and Boltzmann machine.

Pruning has been widely used to slim convolutional neural network (CNN) models to achieve a good trade-off between accuracy and model size so that the pruned models become feasible for power-constrained devices such as mobile phones. This process can be automated to avoid the expensive hand-crafted efforts and to explore a large pruning space automatically so that the high-performance pruning policy can be achieved efficiently. Nowadays, reinforcement learning (RL) and Bayesian optimization (BO)-based auto pruners are widely used due to their solid theoretical foundation, universality, and high compressing quality. However, the RL agent suffers from long training times and high variance of results, while the BO agent is time-consuming for high-dimensional design spaces. In this work, we propose an enhanced BO agent to obtain significant acceleration for auto pruning in high-dimensional design spaces. To achieve this, a novel clustering algorithm is proposed to reduce the dimension of the design space to speedup the searching process. Then, a roll-back algorithm is proposed to recover the high-dimensional design space so that higher pruning accuracy can be obtained. We validate our proposed method on ResNet, MobileNet, and VGG models, and our experiments show that the proposed method significantly improves the accuracy of BO when pruning very deep CNN models. Moreover, our method achieves lower variance and shorter time than the RL-based counterpart.

In recent years, control under urban intersection scenarios becomes an emerging research topic. In such scenarios, the autonomous vehicle confronts complicated situations since it must deal with the interaction with social vehicles timely while obeying the traffic rules. Generally, the autonomous vehicle is supposed to avoid collisions while pursuing better efficiency. The existing work fails to provide a framework that emphasizes the integrity of the scenarios while being able to deploy and test reinforcement learning(RL) methods. Specifically, we propose a benchmark for training and testing RL-based autonomous driving agents in complex intersection scenarios, which is called RL-CIS. Then, a set of baselines are deployed consists of various algorithms. The test benchmark and baselines are to provide a fair and comprehensive training and testing platform for the study of RL for autonomous driving in the intersection scenario, advancing the progress of RL-based methods for intersection autonomous driving control. The code of our proposed framework can be found at https://github.com/liuyuqi123/ComplexUrbanScenarios.

Ensemble reinforcement learning (RL) aims to mitigate instability in Q-learning and to learn a robust policy, which introduces multiple value and policy functions. In this paper, we consider finding a novel but simple ensemble Deep RL algorithm to solve the resource consumption issue. Specifically, we consider integrating multiple models into a single model. To this end, we propose the \underline{M}inimalist \underline{E}nsemble \underline{P}olicy \underline{G}radient framework (MEPG), which introduces minimalist ensemble consistent Bellman update. And we find one value network is sufficient in our framework. Moreover, we theoretically show that the policy evaluation phase in the MEPG is mathematically equivalent to a deep Gaussian Process. To verify the effectiveness of the MEPG framework, we conduct experiments on the gym simulator, which show that the MEPG framework matches or outperforms the state-of-the-art ensemble methods and model-free methods without additional computational resource costs.

Many approaches have been proposed for early classification of time series in light of itssignificance in a wide range of applications including healthcare, transportation and fi-nance. Until now, the early classification problem has been dealt with by considering onlyirrevocable decisions. This paper introduces a new problem calledearly and revocabletimeseries classification, where the decision maker can revoke its earlier decisions based on thenew available measurements. In order to formalize and tackle this problem, we propose anew cost-based framework and derive two new approaches from it. The first approach doesnot consider explicitly the cost of changing decision, while the second one does. Exten-sive experiments are conducted to evaluate these approaches on a large benchmark of realdatasets. The empirical results obtained convincingly show (i) that the ability of revok-ing decisions significantly improves performance over the irrevocable regime, and (ii) thattaking into account the cost of changing decision brings even better results in general.Keywords:revocable decisions, cost estimation, online decision making

We present the single track road problem. In this problem two agents face each-other at opposite positions of a road that can only have one agent pass at a time. We focus on the scenario in which one agent is human, while the other is an autonomous agent. We run experiments with human subjects in a simple grid domain, which simulates the single track road problem. We show that when data is limited, building an accurate human model is very challenging, and that a reinforcement learning agent, which is based on this data, does not perform well in practice. However, we show that an agent that tries to maximize a linear combination of the human's utility and its own utility, achieves a high score, and significantly outperforms other baselines, including an agent that tries to maximize only its own utility. While humans can cope with new situations quite easily, even state-of-the-art algorithms trouble with new situations that they haven't been trained on. Unfortunately, when it comes to autonomous vehicles the results may be devastating. One example for an uncommon, yet important scenario for autonomous vehicles is the problem of a single track road. In this problem two vehicles in opposite directions must cross a narrow road, which is not wide enough to allow both vehicles to pass at the same time.