The idea of experience sharing between cooperative agents naturally emerges from our understanding of how humans learn. Our evolution as a species is tightly linked to the ability to exchange learned knowledge with one another. It follows that experience sharing (ES) between autonomous and independent agents could become the key to accelerate learning in cooperative multiagent settings. We investigate if randomly selecting experiences to share can increase the performance of deep reinforcement learning agents, and propose three new methods for selecting experiences to accelerate the learning process. Firstly, we introduce Focused ES, which prioritizes unexplored regions of the state space. Secondly, we present Prioritized ES, in which temporal-difference error is used as a measure of priority. Finally, we devise Focused Prioritized ES, which combines both previous approaches. The methods are empirically validated in a control problem. While sharing randomly selected experiences between two Deep Q-Network agents shows no improvement over a single agent baseline, we show that the proposed ES methods can successfully outperform the baseline. In particular, the Focused ES accelerates learning by a factor of 2, reducing by 51% the number of episodes required to complete the task.

Knowledge graph embedding, which projects symbolic entities and relations into continuous vector spaces, is gaining increasing attention. Previous methods allow a single static embedding for each entity or relation, ignoring their intrinsic contextual nature, i.e., entities and relations may appear in different graph contexts, and accordingly, exhibit different properties. This work presents Contextualized Knowledge Graph Embedding (CoKE), a novel paradigm that takes into account such contextual nature, and learns dynamic, flexible, and fully contextualized entity and relation embeddings. Two types of graph contexts are studied: edges and paths, both formulated as sequences of entities and relations. CoKE takes a sequence as input and uses a Transformer encoder to obtain contextualized representations. These representations are hence naturally adaptive to the input, capturing contextual meanings of entities and relations therein. Evaluation on a wide variety of public benchmarks verifies the superiority of CoKE in link prediction and path query answering. It performs consistently better than, or at least equally well as current state-of-the-art in almost every case, in particular offering an absolute improvement of 19.7% in H@10 on path query answering. Our code is available at \url{https://github.com/paddlepaddle/models/tree/develop/PaddleKG/CoKE}.

Many reinforcement learning (RL) tasks have specific properties that can be leveraged to modify existing RL algorithms to adapt to those tasks and further improve performance, and a general class of such properties is the multiple reward channel. In those environments the full reward can be decomposed into sub-rewards obtained from different channels. Existing work on reward decomposition either requires prior knowledge of the environment to decompose the full reward, or decomposes reward without prior knowledge but with degraded performance. In this paper, we propose Distributional Reward Decomposition for Reinforcement Learning (DRDRL), a novel reward decomposition algorithm which captures the multiple reward channel structure under distributional setting. Empirically, our method captures the multi-channel structure and discovers meaningful reward decomposition, without any requirements on prior knowledge. Consequently, our agent achieves better performance than existing methods on environments with multiple reward channels.

Traditional first-order logic (FOL) reasoning systems usually rely on manual heuristics for proof guidance. We propose TRAIL: a system that learns to perform proof guidance using reinforcement learning. A key design principle of our system is that it is general enough to allow transfer to problems in different domains that do not share the same vocabulary of the training set. To do so, we developed a novel representation of the internal state of a prover in terms of clauses and inference actions, and a novel neural-based attention mechanism to learn interactions between clauses. We demonstrate that this approach enables the system to generalize from training to test data across domains with different vocabularies, suggesting that the neural architecture in TRAIL is well suited for representing and processing of logical formalisms.

Even though the topic of explainable AI/ML is very popular in text and computer vision domain, most of the previous literatures are not suitable for explaining black-box models' predictions on general data mining datasets. This is because these datasets are usually in high-dimensional vectored features format that are not as friendly and comprehensible as texts and images to the end users. In this paper, we combine the best of both worlds: "explanations by intervention" from causality and "explanations are contrastive" from philosophy and social science domain to explain neural models' predictions for tabular datasets. Specifically, given a model's prediction as label X, we propose a novel idea to intervene and generate minimally modified contrastive sample to be classified as Y, that then results in a simple natural text giving answer to the question "Why X rather than Y?". We carry out experiments with several datasets of different scales and compare our approach with other baselines on three different areas: fidelity, reasonableness and explainability.

Research in multi-agent cooperation has shown that artificial agents are able to learn to play a simple referential game while developing a shared lexicon. This lexicon is not easy to analyze, as it does not show many properties of a natural language. In a simple referential game with two neural network-based agents, we analyze the object-symbol mapping trying to understand what kind of strategy was used to develop the emergent language. We see that, when the environment is uniformly distributed, the agents rely on a random subset of features to describe the objects. When we modify the objects making one feature non-uniformly distributed,the agents realize it is less informative and start to ignore it, and, surprisingly, they make a better use of the remaining features. This interesting result suggests that more natural, less uniformly distributed environments might aid in spurring the emergence of better-behaved languages.

Chris Welty, Praveen Paritosh, Lora Aroyo Google Research Abstract In this paper we present the first steps towards hardening the science of measuring AI systems, by adopting metrology, the science of measurement and its application, and applying it to human (crowd) powered evaluations. We begin with the intuitive observation that evaluating the performance of an AI system is a form of measurement. In all other science and engineering disciplines, the devices used to measure are called instruments, and all measurements are recorded with respect to the characteristics of the instruments used. One does not report mass, speed, or length, for example, of a studied object without disclosing the precision (measurement variance) and resolution (smallest detectable change) of the instrument used. It is extremely common in the AI literature to compare the performance of two systems by using a crowd-sourced dataset as an instrument, but failing to report if the performance difference lies within the capability of that instrument to measure. To illustrate the adoption of metrology to benchmark datasets we use the word similarity benchmark WS353 and several previously published experiments that use it for evaluation. 1 Contributions of this paper In this paper we examine the question of how the variations in human interpretation and other aspects of data collection can affect the measurements we make with crowd-powered datasets. For this, we adopt metrology, the science of measurement and its application, and apply it to human (crowd) powered evaluations. We begin with the intuitive observation that evaluating the performance of an AI system is a form of measurement. In all other science and engineering disciplines, the devices used to measure are called instruments, and all measurements are recorded with respect to the characteristics of the instruments used. One does not report mass, speed, or length, for example, of a studied object without disclosing the precision (measurement variance) and sensitivity (smallest detectable change) of the instrument used. This dataset has been cited over 1500 times, and has spurred the development and evaluation of automated approaches to computing lexical/semantic similarity (Witten and Milne 2008; Agirre et al. 2009) and word embeddings (Mitchell and Lapata 2008; Mikolov et al. 2013; Levy, Goldberg, and Dagan 2015; Pennington, Socher, and Manning 2014; Bojanowski et al. 2017).

The goal of coordinated multi-robot exploration tasks is to employ a team of autonomous robots to explore an unknown environment as quickly as possible. Compared with human-designed methods, which began with heuristic and rule-based approaches, learning-based methods enable individual robots to learn sophisticated and hard-to-design cooperation strategies through deep reinforcement learning technologies. However, in decentralized multi-robot exploration tasks, learning-based algorithms are still far from being universally applicable to the continuous space due to the difficulties associated with area calculation and reward function designing; moreover, existing learning-based methods encounter problems when attempting to balance the historical trajectory issue and target area conflict problem. Furthermore, the scalability of these methods to a large number of agents is poor because of the exponential explosion problem of state space. Accordingly, this paper proposes a novel approach - Multi-head Attention-based Multi-robot Exploration in Continuous Space (MAMECS) - aimed at reducing the state space and automatically learning the cooperation strategies required for decentralized multi-robot exploration tasks in continuous space. Computational geometry knowledge is applied to describe the environment in continuous space and to design an improved reward function to ensure a superior exploration rate. Moreover, the multi-head attention mechanism employed helps to solve the historical trajectory issue in the decentralized multi-robot exploration task, as well as to reduce the quadratic increase of action space.

Slot Filling is the task of extracting the semantic concept from a given natural language utterance. Recently it has been shown that using contextual information, either in work representations (e.g., BERT embedding) or in the computation graph of the model, could improve the performance of the model. However, recent work uses the contextual information in a restricted manner, e.g., by concatenating the word representation and its context feature vector, limiting the model from learning any direct association between the context and the label of word. W e introduce a new deep model utilizing the contextual information for each work in the given sentence in a multi-task setting. Our model enforce consistency between the feature vectors of the context and the word while increasing the expressiveness of the context about the label of the word. Our empirical analysis on a slot filling dataset proves the superiority of the model over the baselines.

This paper presents a computational framework for providing affective labels to real-life situations, called A-Situ. We first define an affective situation, as a specific arrangement of affective entities relevant to emotion elicitation in a situation. Then, the affective situation is represented as a set of labels in the valence-arousal emotion space. Based on physiological behaviors in response to a situation, the proposed framework quantifies the expected emotion evoked by the interaction with a stimulus event. The accumulated result in a spatiotemporal situation is represented as a polynomial curve called the affective curve, which bridges the semantic gap between cognitive and affective perception in real-world situations. We show the efficacy of the curve for reliable emotion labeling in real-world experiments, respectively concerning 1) a comparison between the results from our system and existing explicit assessments for measuring emotion, 2) physiological distinctiveness in emotional states, and 3) physiological characteristics correlated to continuous labels. The efficiency of affective curves to discriminate emotional states is evaluated through subject-dependent classification performance using bicoherence features to represent discrete affective states in the valence-arousal space. Furthermore, electroencephalography-based statistical analysis revealed the physiological correlates of the affective curves.