Equilibrium selection in multi-agent games refers to the problem of selecting a Pareto-optimal equilibrium. It has been shown that many state-of-the-art multi-agent reinforcement learning (MARL) algorithms are prone to converging to Pareto-dominated equilibria due to the uncertainty each agent has about the policy of the other agents during training. To address suboptimal equilibrium selection, we propose Pareto-AC (PAC), an actor-critic algorithm that utilises a simple principle of no-conflict games (a superset of cooperative games with identical rewards): each agent can assume the others will choose actions that will lead to a Pareto-optimal equilibrium. We evaluate PAC in a diverse set of multi-agent games and show that it converges to higher episodic returns compared to alternative MARL algorithms, as well as successfully converging to a Pareto-optimal equilibrium in a range of matrix games. Finally, we propose a graph neural network extension which is shown to efficiently scale in games with up to 15 agents.
The development of autonomous agents which can interact with other agents to accomplish a given task is a core area of research in artificial intelligence and machine learning. Towards this goal, the Autonomous Agents Research Group develops novel machine learning algorithms for autonomous systems control, with a specific focus on deep reinforcement learning and multi-agent reinforcement learning. Research problems include scalable learning of coordinated agent policies and inter-agent communication; reasoning about the behaviours, goals, and composition of other agents from limited observations; and sample-efficient learning based on intrinsic motivation, curriculum learning, causal inference, and representation learning. This article provides a broad overview of the ongoing research portfolio of the group and discusses open problems for future directions.
Achieving safe and robust autonomy is the key bottleneck on the path towards broader adoption of autonomous vehicles technology. This motivates going beyond extrinsic metrics such as miles between disengagement, and calls for approaches that embody safety by design. In this paper, we address some aspects of this challenge, with emphasis on issues of motion planning and prediction. We do this through description of novel approaches taken to solving selected sub-problems within an autonomous driving stack, in the process introducing the design philosophy being adopted within Five. This includes safe-by-design planning, interpretable as well as verifiable prediction, and modelling of perception errors to enable effective sim-to-real and real-to-sim transfer within the testing pipeline of a realistic autonomous system.
Ad hoc teamwork (AHT) is the problem of creating an agent that must collaborate with previously unseen teammates without prior coordination. Many existing AHT methods can be categorised as type-based methods, which require a set of predefined teammates for training. Designing teammate types for training is a challenging issue that determines the generalisation performance of agents when dealing with teammate types unseen during training. In this work, we propose a method to discover diverse teammate types based on maximising best response diversity metrics. We show that our proposed approach yields teammate types that require a wider range of best responses from the learner during collaboration, which potentially improves the robustness of a learner's performance in AHT compared to alternative methods.
We propose the novel few-shot teamwork (FST) problem, where skilled agents trained in a team to complete one task are combined with skilled agents from different tasks, and together must learn to adapt to an unseen but related task. We discuss how the FST problem can be seen as addressing two separate problems: one of reducing the experience required to train a team of agents to complete a complex task; and one of collaborating with unfamiliar teammates to complete a new task. Progress towards solving FST could lead to progress in both multi-agent reinforcement learning and ad hoc teamwork.
In real-world robotics applications, Reinforcement Learning (RL) agents are often unable to generalise to environment variations that were not observed during training. This issue is intensified for image-based RL where a change in one variable, such as the background colour, can change many pixels in the image, and in turn can change all values in the agent's internal representation of the image. To learn more robust representations, we introduce TEmporal Disentanglement (TED), a self-supervised auxiliary task that leads to disentangled representations using the sequential nature of RL observations. We find empirically that RL algorithms with TED as an auxiliary task adapt more quickly to changes in environment variables with continued training compared to state-of-the-art representation learning methods. Due to the disentangled structure of the representation, we also find that policies trained with TED generalise better to unseen values of variables irrelevant to the task (e.g. background colour) as well as unseen values of variables that affect the optimal policy (e.g. goal positions).
Successful deployment of multi-agent reinforcement learning often requires agents to adapt their behaviour. In this work, we discuss the problem of teamwork adaptation in which a team of agents needs to adapt their policies to solve novel tasks with limited fine-tuning. Motivated by the intuition that agents need to be able to identify and distinguish tasks in order to adapt their behaviour to the current task, we propose to learn multi-agent task embeddings (MATE). These task embeddings are trained using an encoder-decoder architecture optimised for reconstruction of the transition and reward functions which uniquely identify tasks. We show that a team of agents is able to adapt to novel tasks when provided with task embeddings. We propose three MATE training paradigms: independent MATE, centralised MATE, and mixed MATE which vary in the information used for the task encoding. We show that the embeddings learned by MATE identify tasks and provide useful information which agents leverage during adaptation to novel tasks.
When used in autonomous driving, goal recognition allows the future behaviour of other vehicles to be more accurately predicted. A recent goal recognition method for autonomous vehicles, GRIT, has been shown to be fast, accurate, interpretable and verifiable. In autonomous driving, vehicles can encounter novel scenarios that were unseen during training, and the environment is partially observable due to occlusions. However, GRIT can only operate in fixed frame scenarios, with full observability. We present a novel goal recognition method named Goal Recognition with Interpretable Trees under Occlusion (OGRIT), which solves these shortcomings of GRIT. We demonstrate that OGRIT can generalise between different scenarios and handle missing data due to occlusions, while still being fast, accurate, interpretable and verifiable.
Inscrutable AI systems are difficult to trust, especially if they operate in safety-critical settings like autonomous driving. Therefore, there is a need to build transparent and queryable systems to increase trust levels. We propose a transparent, human-centric explanation generation method for autonomous vehicle motion planning and prediction based on an existing white-box system called IGP2. Our method integrates Bayesian networks with context-free generative rules and can give causal natural language explanations for the high-level driving behaviour of autonomous vehicles. Preliminary testing on simulated scenarios shows that our method captures the causes behind the actions of autonomous vehicles and generates intelligible explanations with varying complexity.
Learning control from pixels is difficult for reinforcement learning (RL) agents because representation learning and policy learning are intertwined. Previous approaches remedy this issue with auxiliary representation learning tasks, but they either do not consider the temporal aspect of the problem or only consider single-step transitions. Instead, we propose Hierarchical $k$-Step Latent (HKSL), an auxiliary task that learns representations via a hierarchy of forward models that operate at varying magnitudes of step skipping while also learning to communicate between levels in the hierarchy. We evaluate HKSL in a suite of 30 robotic control tasks and find that HKSL either reaches higher episodic returns or converges to maximum performance more quickly than several current baselines. Also, we find that levels in HKSL's hierarchy can learn to specialize in long- or short-term consequences of agent actions, thereby providing the downstream control policy with more informative representations. Finally, we determine that communication channels between hierarchy levels organize information based on both sides of the communication process, which improves sample efficiency.