Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
originality: Moderate, the task have being studied in previous negotiation competition like D-brane, but the SL+ RL approach to study this problem is new. quality: Experiments are very solid, results are well presented. However, more theoretical analysis of the problem and intellectual insight would hugely improve the work. clarity: Well organized writing. However, the paper definitely does not provide enough information for possibility to reproduce the results. significance: Moderate. The approach that by collecting a large dataset of human games and do SL on it, later improve with RL is not new, but the execution of this idea on diplomacy is still a non-trivial job. This paper looks like a well-written, well-executed project report to me, though it should give either more insights, theories or more detailed algorithm, dataset and code.
The dynamically changing alliances mean that the domain of diplomacy presents unique challenges for agents. I agree with the authors that this means that diplomacy is ‘deserving of special attention’, I would consider the full game to be a grand challenge for multi-agent research. With recent progress in large-scale RL focusing on single-agent and 2-player zero sum games, this problem is particularly timely. This work presents state of the art agents trained with deep learning. To my knowledge this is the first successful application of deep learning to diplomacy. The design of the neural network is given particular attention: Several of the design choices are studied by ablation in the supervised task. The dataset of games is large, I’m not sure how many online games of diplomacy have been played, but 150,000 seems likely to be a significant proportion of them. Not all the games are No-Press, so there is a domain mis-match for much of the dataset. Maybe performance could be improved by including press rules as a feature for the network, or training more on the no-press games. I would like to see more details on how the dataset was selected. Sometimes in online games players fail to enter orders, or cheat by creating multiple accounts. Have any preprocessing steps been taken to remove data where these things have occurred? To evaluate the performance of agents, the SL DipNet played games against 4 rule-based baselines, including previous SOTA, plus the RL DipNet, under the conditions of 1 SL Dipnet vs 6 of the opponent, and 1 of the opponent vs 6 SL DipNet. Trueskill is then calculated based on these games. These 1 vs 6 matches are interesting to see, and clearly demonstrate that the DipNet programs outperform the other agents. However, we _cannot_ draw conclusions about the relative strength of the RL and SL agents from this tournament. The RL agent has had access to the SL agent to train against by the warm-start procedure, and the experiment never compares the RL DipNet against the baselines. I’d like the Trueskill to be generated by some more varied structure of games: A simple procedure that would be interesting is where each game, the player to play each power is sampled iid uniformly from the 6 different agents. Then we could understand how the agents perform against varied and heterogeneous opponents. This would test an important question in diplomacy, of whether self-play RL in 3+ agent problems overfits to its own play-style. The coalition analysis is nice, and I am particularly interested in the analysis around the effect of RL training on effective cross power support. One point to note is that cross power support in human no-press is likely to be reduced by virtue of signalling actions, where a move that is guaranteed to succeed is nonetheless supported by another player to signal that they want to ally, and help with coordination. I wonder if a metric for this behaviour can be designed, whether the SL network is able to pick up on such signalling moves, and whether RL changes this ability? Minor Comments: Line 28: Players can have a maximum of 17 units, because if they have 18+ units, they have already won the game. How was the average of 26 possible actions per unit calculated? Is that the empirical average in a dataset after removing any support/convoy orders that use non-existent units? Line 135: Did this mean to refer to  Learning with Opponent Learning Awareness, rather than the DIAL paper? Line 175: “there is no weight sharing”: Does this mean that there is no sharing of weights for the convolutions at different nodes of the graph, or just none between different layers? Would be good to clarify. ======= Comments following Author Response: Thanks to the authors for their clarification on how Trueskill was calculated. This makes it much easier to understand what the scores mean. There's some additional information I'd like to see included in the final paper if possible: - Table of results - Gameplay data (i.e. example games) - Report the standard deviations of Trueskill estimates
Comments after rebuttal and discussion ====================================== Thank you for your rebuttal. It appears you have foreseen many of the concerns and there simply isn't a great solution for many of them. It may be helpful to note in the body of the paper that the particular evaluation suffices here because the skill gap is so large in the agent pool. In the future as the pool of agents gets stronger a different evaluation may be necessary. Original Review =============== This paper is difficult to evaluate as a whole. I lean towards acceptance, but primarily because this paper may serve as a catalyst for future work. DipNet and its evaluation are not as thorough as one would hope for. The game of Diplomacy is well described and motivated as a challenging multi-agent AI domain. It is clear that it possess interesting aspects that are not represented in popular games in the literature---specifically the competitive/cooperative aspect as well as its communication structure. This paper is not the first to consider diplomacy as the authors point out, but past work, like Shapiro et al, was perhaps too early to spur academic interest. DipNet itself is non-trivial, but it is not incredibly insightful on its own. e.g., the state representation is domain-specific and the LSTM action decoding is heuristic. Bootstrapping training from human data and tuning with self-play is common practice. i.e., there are many seemingly important details, none of which are novel. The quality of this paper's contribution is mixed. On one hand, the motivation and exposition of Diplomacy as a test domain are excellent. The development and evaluation of DipNet is lacking, though. As mentioned above, the DipNet agent is fairly complicated, i.e., numerous non-trivial decisions were involved in its creation, many of which are parametrized. Typically, these decisions are not thoroughly justified and it is not clear how reliant the agent's performance is on these decisions. e.g., the state representation is lossy. The authors state "Based on our empirical findings, orders from the last movement phase are enough to infer the current relationship between the powers", but no evidence or procedure to reproduce this claim is provided. Similarly, using the no-check version of the game is claimed to be important to enable communication. This seems intuitive, but no evidence is provided to show this is indeed true or to validate the magnitude of this. Other decisions, such as the network structure, LSTM action decoding order, training procedure and parameters, reward and reward shaping, require further details and justification to aid in reproducibility. Perhaps the most important criticism of the paper is its evaluation of DipNet. The 1 vs. 6 head-to-head play demonstrates that DipNet is much stronger than the baselines, but the same approach is likely to fail when evaluating agents that are closer in strength. In multi-agent scenarios, is often the case that an agent trained with self-play will learn to implicitly collude with itself. e.g., if an agent indeed finds a Nash equilibrium then the solo agent will be at a tremendous disadvantage. This can even occur with different agents that trained in a similar fashion, e.g., playing 3 copies of agent A vs. 2 copies of agent B vs. 2 copies of agent C may not be informative if A and B are similar. i.e., using overall utility or games won as a performance indicator is incredibly sensitive to the pool of agents, especially in multi-agent scenarios. A thorough and thoughtful evaluation is of particular importance here as it sets a precedent for future papers. Again, I appreciate that the authors are willing to release their source code as it enables others to perform a different evaluation in the future should a more suitable one come to light. Minor comments: 269: metod => method Please go through the citations to correct capitalization, e.g.,  nature => Nature