Monte Carlo Tree Search (MCTS) is an immensely popular search-based framework used for decision making. It is traditionally applied to domains where a perfect simulation model of the environment is available. We study and improve MCTS in the context where the environment model is given but imperfect. We show that the discrepancy between the model and the actual environment can lead to significant performance degradation with standard MCTS. We therefore develop Uncertainty Adapted MCTS (UA-MCTS), a more robust algorithm within the MCTS framework. We estimate the transition uncertainty in the given model, and direct the search towards more certain transitions in the state space. We modify all four MCTS phases to improve the search behavior by considering these estimates. We prove, in the corrupted bandit case, that adding uncertainty information to adapt UCB leads to tighter regret bound than standard UCB. Empirically, we evaluate UA-MCTS and its individual components on the deterministic domains from the MinAtar test suite. Our results demonstrate that UA-MCTS strongly improves MCTS in the presence of model transition errors.
In this work, we adapt a training approach inspired by the original AlphaGo system to play the imperfect information game of Reconnaissance Blind Chess. Using only the observations instead of a full description of the game state, we first train a supervised agent on publicly available game records. Next, we increase the performance of the agent through self-play with the on-policy reinforcement learning algorithm Proximal Policy Optimization. We do not use any search to avoid problems caused by the partial observability of game states and only use the policy network to generate moves when playing. With this approach, we achieve an ELO of 1330 on the RBC leaderboard, which places our agent at position 27 at the time of this writing. We see that self-play significantly improves performance and that the agent plays acceptably well without search and without making assumptions about the true game state.
In this paper, we study learning in probabilistic domains where the learner may receive incorrect labels but can improve the reliability of labels by repeatedly sampling them. In such a setting, one faces the problem of whether the fixed budget for obtaining training examples should rather be used for obtaining all different examples or for improving the label quality of a smaller number of examples by re-sampling their labels. We motivate this problem in an application to compare the strength of poker hands where the training signal depends on the hidden community cards, and then study it in depth in an artificial setting where we insert controlled noise levels into the MNIST database. Our results show that with increasing levels of noise, resampling previous examples becomes increasingly more important than obtaining new examples, as classifier performance deteriorates when the number of incorrect labels is too high. In addition, we propose two different validation strategies; switching from lower to higher validations over the course of training and using chi-square statistics to approximate the confidence in obtained labels.
Scaling up the size and training of autoregressive language models has enabled novel ways of solving Natural Language Processing tasks using zero-shot and few-shot learning. While extreme-scale language models such as GPT-3 offer multilingual capabilities, zero-shot learning for languages other than English remain largely unexplored. Here, we introduce Cedille, a large open source auto-regressive language model, specifically trained for the French language. Our results show that Cedille outperforms existing French language models and is competitive with GPT-3 on a range of French zero-shot benchmarks. Furthermore, we provide an in-depth comparison of the toxicity exhibited by these models, showing that Cedille marks an improvement in language model safety thanks to dataset filtering.
Abnormal states in deep reinforcement learning~(RL) are states that are beyond the scope of an RL policy. Such states may make the RL system unsafe and impede its deployment in real scenarios. In this paper, we propose a simple yet effective anomaly detection framework for deep RL algorithms that simultaneously considers random, adversarial and out-of-distribution~(OOD) state outliers. In particular, we attain the class-conditional distributions for each action class under the Gaussian assumption, and rely on these distributions to discriminate between inliers and outliers based on Mahalanobis Distance~(MD) and Robust Mahalanobis Distance. We conduct extensive experiments on Atari games that verify the effectiveness of our detection strategies. To the best of our knowledge, we present the first in-detail study of statistical and adversarial anomaly detection in deep RL algorithms. This simple unified anomaly detection paves the way towards deploying safe RL systems in real-world applications.
Materials' microstructures are signatures of their alloying composition and processing history. Therefore, microstructures exist in a wide variety. As materials become increasingly complex to comply with engineering demands, advanced computer vision (CV) approaches such as deep learning (DL) inevitably gain relevance for quantifying microstrucutures' constituents from micrographs. While DL can outperform classical CV techniques for many tasks, shortcomings are poor data efficiency and generalizability across datasets. This is inherently in conflict with the expense associated with annotating materials data through experts and extensive materials diversity. To tackle poor domain generalizability and the lack of labeled data simultaneously, we propose to apply a sub-class of transfer learning methods called unsupervised domain adaptation (UDA). These algorithms address the task of finding domain-invariant features when supplied with annotated source data and unannotated target data, such that performance on the latter distribution is optimized despite the absence of annotations. Exemplarily, this study is conducted on a lath-shaped bainite segmentation task in complex phase steel micrographs. Here, the domains to bridge are selected to be different metallographic specimen preparations (surface etchings) and distinct imaging modalities. We show that a state-of-the-art UDA approach surpasses the na\"ive application of source domain trained models on the target domain (generalization baseline) to a large extent. This holds true independent of the domain shift, despite using little data, and even when the baseline models were pre-trained or employed data augmentation. Through UDA, mIoU was improved over generalization baselines from 82.2%, 61.0%, 49.7% to 84.7%, 67.3%, 73.3% on three target datasets, respectively. This underlines this techniques' potential to cope with materials variance.
In this paper, we study the problem of evaluating the addition of elements to a set. This problem is difficult, because it can, in the general case, not be reduced to unconditional preferences between the choices. Therefore, we model preferences based on the context of the decision. We discuss and compare two different Siamese network architectures for this task: a twin network that compares the two sets resulting after the addition, and a triplet network that models the contribution of each candidate to the existing set. We evaluate the two settings on a real-world task; learning human card preferences for deck building in the collectible card game Magic: The Gathering. We show that the triplet approach achieves a better result than the twin network and that both outperform previous results on this task.
Drafting, i.e., the selection of a subset of items from a larger candidate set, is a key element of many games and related problems. It encompasses team formation in sports or e-sports, as well as deck selection in many modern card games. The key difficulty of drafting is that it is typically not sufficient to simply evaluate each item in a vacuum and to select the best items. The evaluation of an item depends on the context of the set of items that were already selected earlier, as the value of a set is not just the sum of the values of its members - it must include a notion of how well items go together. In this paper, we study drafting in the context of the card game Magic: The Gathering. We propose the use of a contextual preference network, which learns to compare two possible extensions of a given deck of cards. We demonstrate that the resulting network is better able to evaluate card decks in this game than previous attempts.
Boolean Satisfiability (SAT) is a well-known NP-complete problem. Despite this theoretical hardness, SAT solvers based on Conflict Driven Clause Learning (CDCL) can solve large SAT instances from many important domains. CDCL learns clauses from conflicts, a technique that allows a solver to prune its search space. The selection heuristics in CDCL prioritize variables that are involved in recent conflicts. While only a fraction of decisions generate any conflicts, many generate multiple conflicts. In this paper, we study conflict-generating decisions in CDCL in detail. We investigate the impact of single conflict (sc) decisions, which generate only one conflict, and multi-conflict (mc) decisions which generate two or more. We empirically characterize these two types of decisions based on the quality of the learned clauses produced by each type of decision. We also show an important connection between consecutive clauses learned within the same mc decision, where one learned clause triggers the learning of the next one forming a chain of clauses. This leads to the consideration of similarity between conflicts, for which we formulate the notion of conflictsproximity as a similarity measure. We show that conflicts in mc decisions are more closely related than consecutive conflicts generated from sc decisions. Finally, we develop Common Reason Variable Reduction (CRVR) as a new decision strategy that reduces the selection priority of some variables from the learned clauses of mc decisions. Our empirical evaluation of CRVR implemented in three leading solvers demonstrates performance gains in benchmarks from the main track of SAT Competition-2020.
Social media analysis has become a common approach to assess public opinion on various topics, including those about health, in near real-time. The growing volume of social media posts has led to an increased usage of modern machine learning methods in natural language processing. While the rapid dynamics of social media can capture underlying trends quickly, it also poses a technical problem: algorithms trained on annotated data in the past may underperform when applied to contemporary data. This phenomenon, known as concept drift, can be particularly problematic when rapid shifts occur either in the topic of interest itself, or in the way the topic is discussed. Here, we explore the effect of machine learning concept drift by focussing on vaccine sentiments expressed on Twitter, a topic of central importance especially during the COVID-19 pandemic. We show that while vaccine sentiment has declined considerably during the COVID-19 pandemic in 2020, algorithms trained on pre-pandemic data would have largely missed this decline due to concept drift. Our results suggest that social media analysis systems must address concept drift in a continuous fashion in order to avoid the risk of systematic misclassification of data, which is particularly likely during a crisis when the underlying data can change suddenly and rapidly.