ETGL-DDPG: A Deep Deterministic Policy Gradient Algorithm for Sparse Reward Continuous Control

We consider deep deterministic policy gradient (DDPG) in the context of reinforcement learning with sparse rewards. To enhance exploration, we introduce a search procedure, \emph{${\epsilon}{t}$-greedy}, which generates exploratory options for exploring less-visited states. We prove that search using $\epsilon t$-greedy has polynomial sample complexity under mild MDP assumptions. To more efficiently use the information provided by rewarded transitions, we develop a new dual experience replay buffer framework, \emph{GDRB}, and implement \emph{longest n-step returns}. The resulting algorithm, \emph{ETGL-DDPG}, integrates all three techniques: \bm{$\epsilon t$}-greedy, \textbf{G}DRB, and \textbf{L}ongest $n$-step, into DDPG. We evaluate ETGL-DDPG on standard benchmarks and demonstrate that it outperforms DDPG, as well as other state-of-the-art methods, across all tested sparse-reward continuous environments. Ablation studies further highlight how each strategy individually enhances the performance of DDPG in this setting.

Efficiently Training Neural Networks for Imperfect Information Games by Sampling Information Sets

In imperfect information games, the evaluation of a game state not only depends on the observable world but also relies on hidden parts of the environment. As accessing the obstructed information trivialises state evaluations, one approach to tackle such problems is to estimate the value of the imperfect state as a combination of all states in the information set, i.e., all possible states that are consistent with the current imperfect information. In this work, the goal is to learn a function that maps from the imperfect game information state to its expected value. However, constructing a perfect training set, i.e. an enumeration of the whole information set for numerous imperfect states, is often infeasible. To compute the expected values for an imperfect information game like \textit{Reconnaissance Blind Chess}, one would need to evaluate thousands of chess positions just to obtain the training target for a single state. Still, the expected value of a state can already be approximated with appropriate accuracy from a much smaller set of evaluations. Thus, in this paper, we empirically investigate how a budget of perfect information game evaluations should be distributed among training samples to maximise the return. Our results show that sampling a small number of states, in our experiments roughly 3, for a larger number of separate positions is preferable over repeatedly sampling a smaller quantity of states. Thus, we find that in our case, the quantity of different samples seems to be more important than higher target quality.

Neural Network-based Information Set Weighting for Playing Reconnaissance Blind Chess

In imperfect information games, the game state is generally not fully observable to players. Therefore, good gameplay requires policies that deal with the different information that is hidden from each player. To combat this, effective algorithms often reason about information sets; the sets of all possible game states that are consistent with a player's observations. While there is no way to distinguish between the states within an information set, this property does not imply that all states are equally likely to occur in play. We extend previous research on assigning weights to the states in an information set in order to facilitate better gameplay in the imperfect information game of Reconnaissance Blind Chess. For this, we train two different neural networks which estimate the likelihood of each state in an information set from historical game data. Experimentally, we find that a Siamese neural network is able to achieve higher accuracy and is more efficient than a classical convolutional neural network for the given domain. Finally, we evaluate an RBC-playing agent that is based on the generated weightings and compare different parameter settings that influence how strongly it should rely on them. The resulting best player is ranked 5th on the public leaderboard.

Contrastive Learning of Preferences with a Contextual InfoNCE Loss

A common problem in contextual preference ranking is that a single preferred action is compared against several choices, thereby blowing up the complexity and skewing the preference distribution. In this work, we show how one can solve this problem via a suitable adaptation of the CLIP framework.This adaptation is not entirely straight-forward, because although the InfoNCE loss used by CLIP has achieved great success in computer vision and multi-modal domains, its batch-construction technique requires the ability to compare arbitrary items, and is not well-defined if one item has multiple positive associations in the same batch. We empirically demonstrate the utility of our adapted version of the InfoNCE loss in the domain of collectable card games, where we aim to learn an embedding space that captures the associations between single cards and whole card pools based on human selections. Such selection data only exists for restricted choices, thus generating concrete preferences of one item over a set of other items rather than a perfect fit between the card and the pool. Our results show that vanilla CLIP does not perform well due to the aforementioned intuitive issues. However, by adapting CLIP to the problem, we receive a model outperforming previous work trained with the triplet loss, while also alleviating problems associated with mining triplets.

Learning With Generalised Card Representations for "Magic: The Gathering"

A defining feature of collectable card games is the deck building process prior to actual gameplay, in which players form their decks according to some restrictions. Learning to build decks is difficult for players and models alike due to the large card variety and highly complex semantics, as well as requiring meaningful card and deck representations when aiming to utilise AI. In addition, regular releases of new card sets lead to unforeseeable fluctuations in the available card pool, thus affecting possible deck configurations and requiring continuous updates. Previous Game AI approaches to building decks have often been limited to fixed sets of possible cards, which greatly limits their utility in practice. In this work, we explore possible card representations that generalise to unseen cards, thus greatly extending the real-world utility of AI-based deck building for the game "Magic: The Gathering".We study such representations based on numerical, nominal, and text-based features of cards, card images, and meta information about card usage from third-party services. Our results show that while the particular choice of generalised input representation has little effect on learning to predict human card selections among known cards, the performance on new, unseen cards can be greatly improved. Our generalised model is able to predict 55\% of human choices on completely unseen cards, thus showing a deep understanding of card quality and strategy.

Expected Work Search: Combining Win Rate and Proof Size Estimation

We propose Expected Work Search (EWS), a new game solving algorithm. EWS combines win rate estimation, as used in Monte Carlo Tree Search, with proof size estimation, as used in Proof Number Search. The search efficiency of EWS stems from minimizing a novel notion of Expected Work, which predicts the expected computation required to solve a position. EWS outperforms traditional solving algorithms on the games of Go and Hex. For Go, we present the first solution to the empty 5x5 board with the commonly used positional superko ruleset. For Hex, our algorithm solves the empty 8x8 board in under 4 minutes. Experiments show that EWS succeeds both with and without extensive domain-specific knowledge.

Monte Carlo Tree Search in the Presence of Transition Uncertainty

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.

Supervised and Reinforcement Learning from Observations in Reconnaissance Blind Chess

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.

Quantity vs Quality: Investigating the Trade-Off between Sample Size and Label Reliability

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.

Cedille: A large autoregressive French language model

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.