Behavior prediction in dynamic, multi-agent systems is an important problem in the context of self-driving cars, due to the complex representations and interactions of road components, including moving agents (e.g. pedestrians and vehicles) and road context information (e.g. lanes, traffic lights). This paper introduces VectorNet, a hierarchical graph neural network that first exploits the spatial locality of individual road components represented by vectors and then models the high-order interactions among all components. In contrast to most recent approaches, which render trajectories of moving agents and road context information as bird-eye images and encode them with convolutional neural networks (ConvNets), our approach operates on a vector representation. By operating on the vectorized high definition (HD) maps and agent trajectories, we avoid lossy rendering and computationally intensive ConvNet encoding steps. To further boost VectorNet's capability in learning context features, we propose a novel auxiliary task to recover the randomly masked out map entities and agent trajectories based on their context. We evaluate VectorNet on our in-house behavior prediction benchmark and the recently released Argoverse forecasting dataset. Our method achieves on par or better performance than the competitive rendering approach on both benchmarks while saving over 70% of the model parameters with an order of magnitude reduction in FLOPs. It also outperforms the state of the art on the Argoverse dataset.

In the future we can expect that artificial intelligent agents, once deployed, will be required to learn continually from their experience during their operational lifetime. Such agents will also need to communicate with humans and other agents regarding the content of their experience, in the context of passing along their learnings, for the purpose of explaining their actions in specific circumstances or simply to relate more naturally to humans concerning experiences the agent acquires that are not necessarily related to their assigned tasks. We argue that to support these goals, an agent would benefit from an episodic memory; that is, a memory that encodes the agent's experience in such a way that the agent can relive the experience, communicate about it and use its past experience, inclusive of the agents own past actions, to learn more effective models and policies. In this short paper, we propose one potential approach to provide an AI agent with such capabilities. We draw upon the ever-growing body of work examining the function and operation of the Medial Temporal Lobe (MTL) in mammals to guide us in adding an episodic memory capability to an AI agent composed of artificial neural networks (ANNs). Based on that, we highlight important aspects to be considered in the memory organization and we propose an architecture combining ANNs and standard Computer Science techniques for supporting storage and retrieval of episodic memories. Despite being initial work, we hope this short paper can spark discussions around the creation of intelligent agents with memory or, at least, provide a different point of view on the subject.

Theory of Mind is commonly defined as the ability to attribute mental states (e.g., beliefs, goals) to oneself, and to others. A large body of previous work - from the social sciences to artificial intelligence - has observed that Theory of Mind capabilities are central to providing an explanation to another agent or when explaining that agent's behaviour. In this paper, we build and expand upon previous work by providing an account of explanation in terms of the beliefs of agents and the mechanism by which agents revise their beliefs given possible explanations. We further identify a set of desiderata for explanations that utilize Theory of Mind. These desiderata inform our belief-based account of explanation.

Functional decision theory (FDT) is a fairly new mode of decision theory and a normative viewpoint on how an agent should maximize expected utility. The current standard in decision theory and computer science is causal decision theory (CDT), largely seen as superior to the main alternative evidential decision theory (EDT). These theories prescribe three distinct methods for maximizing utility. We explore how FDT differs from CDT and EDT, and what implications it has on the behavior of FDT agents and humans. It has been shown in previous research how FDT can outperform CDT and EDT. We additionally show FDT performing well on more classical game theory problems and argue for its extension to human problems to show that its potential for superiority is robust. We also make FDT more concrete by displaying it in an evolutionary environment, competing directly against other theories.

Dialogue engines that incorporate different types of agents to converse with humans are popular. However, conversations are dynamic in the sense that a selected response will change the conversation on-the-fly, influencing the subsequent utterances in the conversation, which makes the response selection a challenging problem. We model the problem of selecting the best response from a set of responses generated by a heterogeneous set of dialogue agents by taking into account the conversational history, and propose a \emph{Neural Response Selection} method. The proposed method is trained to predict a coherent set of responses within a single conversation, considering its own predictions via a curriculum training mechanism. Our experimental results show that the proposed method can accurately select the most appropriate responses, thereby significantly improving the user experience in dialogue systems.

Neural networks are effective function approximators, but hard to train in the reinforcement learning (RL) context mainly because samples are correlated. For years, scholars have got around this by employing experience replay or an asynchronous parallel-agent system. This paper proposes Discrete-to-Deep Supervised Policy Learning (D2D-SPL) for training neural networks in RL. D2D-SPL discretises the continuous state space into discrete states and uses actor-critic to learn a policy. It then selects from each discrete state an input value and the action with the highest numerical preference as an input/target pair. Finally it uses input/target pairs from all discrete states to train a classifier. D2D-SPL uses a single agent, needs no experience replay and learns much faster than state-of-the-art methods. We test our method with two RL environments, the Cartpole and an aircraft manoeuvring simulator.

Federated learning has emerged as an umbrella term for centralized coordination strategies in multi-agent environments. While many federated learning architectures process data in an online manner, and are hence adaptive by nature, most performance analyses assume static optimization problems and offer no guarantees in the presence of drifts in the problem solution or data characteristics. We consider a federated learning model where at every iteration, a random subset of available agents perform local updates based on their data. Under a non-stationary random walk model on the true minimizer for the aggregate optimization problem, we establish that the performance of the architecture is determined by three factors, namely, the data variability at each agent, the model variability across all agents, and a tracking term that is inversely proportional to the learning rate of the algorithm. The results clarify the trade-off between convergence and tracking performance.

We consider the problem of demand-side energy management, where each household is equipped with a smart meter that is able to schedule home appliances online. The goal is to minimise the overall cost under a real-time pricing scheme. While previous works have introduced centralised approaches, we formulate the smart grid environment as a Markov game, where each household is a decentralised agent, and the grid operator produces a price signal that adapts to the energy demand. The main challenge addressed in our approach is partial observability and perceived non-stationarity of the environment from the viewpoint of each agent. We propose a multi-agent extension of a deep actor-critic algorithm that shows success in learning in this environment. This algorithm learns a centralised critic that coordinates training of all agents. Our approach thus uses centralised learning but decentralised execution. Simulation results show that our online deep reinforcement learning method can reduce both the peak-to-average ratio of total energy consumed and the cost of electricity for all households based purely on instantaneous observations and a price signal.

We consider infinite horizon dynamic programming problems, where the control at each stage consists of several distinct decisions, each one made by one of several agents. In an earlier work we introduced a policy iteration algorithm, where the policy improvement is done one-agent-at-a-time in a given order, with knowledge of the choices of the preceding agents in the order. As a result, the amount of computation for each policy improvement grows linearly with the number of agents, as opposed to exponentially for the standard all-agents-at-once method. For the case of a finite-state discounted problem, we showed convergence to an agent-by-agent optimal policy. In this paper, this result is extended to value iteration and optimistic versions of policy iteration, as well as to more general DP problems where the Bellman operator is a contraction mapping, such as stochastic shortest path problems with all policies being proper.

Systems based on artificial intelligence and machine learning models should be transparent, in the sense of being capable of explaining their decisions to gain humans' approval and trust. While there are a number of explainability techniques that can be used to this end, many of them are only capable of outputting a single one-size-fits-all explanation that simply cannot address all of the explainees' diverse needs. In this work we introduce a model-agnostic and post-hoc local explainability technique for black-box predictions called LIMEtree, which employs surrogate multi-output regression trees. We validate our algorithm on a deep neural network trained for object detection in images and compare it against Local Interpretable Model-agnostic Explanations (LIME). Our method comes with local fidelity guarantees and can produce a range of diverse explanation types, including contrastive and counterfactual explanations praised in the literature. Some of these explanations can be interactively personalised to create bespoke, meaningful and actionable insights into the model's behaviour. While other methods may give an illusion of customisability by wrapping, otherwise static, explanations in an interactive interface, our explanations are truly interactive, in the sense of allowing the user to "interrogate" a black-box model. LIMEtree can therefore produce consistent explanations on which an interactive exploratory process can be built.