Multiagent decision-making in partially observable environments is usually modelled as either an extensive-form game (EFG) in game theory or a partially observable stochastic game (POSG) in multiagent reinforcement learning (MARL). One issue with the current situation is that while most practical problems can be modelled in both formalisms, the relationship of the two models is unclear, which hinders the transfer of ideas between the two communities. A second issue is that while EFGs have recently seen significant algorithmic progress, their classical formalization is unsuitable for efficient presentation of the underlying ideas, such as those around decomposition. To solve the first issue, we introduce factored-observation stochastic games (FOSGs), a minor modification of the POSG formalism which distinguishes between private and public observation and thereby greatly simplifies decomposition. To remedy the second issue, we show that FOSGs and POSGs are naturally connected to EFGs: by "unrolling" a FOSG into its tree form, we obtain an EFG. Conversely, any perfect-recall timeable EFG corresponds to some underlying FOSG in this manner. Moreover, this relationship justifies several minor modifications to the classical EFG formalization that recently appeared as an implicit response to the model's issues with decomposition. Finally, we illustrate the transfer of ideas between EFGs and MARL by presenting three key EFG techniques -- counterfactual regret minimization, sequence form, and decomposition -- in the FOSG framework.

Researchers have shown that neural networks are vulnerable to adversarial examples and subtle environment changes. The resulting errors can look like blunders to humans, eroding trust in these agents. In prior games research, agent evaluation often focused on the in-practice game outcomes. Such evaluation typically fails to evaluate robustness to worst-case outcomes. Computer poker research has examined how to assess such worst-case performance. Unfortunately, exact computation is infeasible with larger domains, and existing approximations are poker-specific. We introduce ISMCTS-BR, a scalable search-based deep reinforcement learning algorithm for learning a best response to an agent, approximating worst-case performance. We demonstrate the technique in several games against a variety of agents, including several AlphaZero-based agents. Supplementary material is available at https://arxiv.org/abs/2004.09677.

For artificially intelligent learning systems to have widespread applicability in real-world settings, it is important that they be able to operate decentrally. Unfortunately, decentralized control is difficult---computing even an epsilon-optimal joint policy is a NEXP complete problem. Nevertheless, a recently rediscovered insight---that a team of agents can coordinate via common knowledge---has given rise to algorithms capable of finding optimal joint policies in small common-payoff games. The Bayesian action decoder (BAD) leverages this insight and deep reinforcement learning to scale to games as large as two-player Hanabi. However, the approximations it uses to do so prevent it from discovering optimal joint policies even in games small enough to brute force optimal solutions. This work proposes CAPI, a novel algorithm which, like BAD, combines common knowledge with deep reinforcement learning. However, unlike BAD, CAPI prioritizes the propensity to discover optimal joint policies over scalability. While this choice precludes CAPI from scaling to games as large as Hanabi, empirical results demonstrate that, on the games to which CAPI does scale, it is capable of discovering optimal joint policies even when other modern multi-agent reinforcement learning algorithms are unable to do so.

In this paper we investigate the Follow the Regularized Leader dynamics in sequential imperfect information games (IIG). We generalize existing results of Poincar{é} recurrence from normal-form games to zero-sum two-player imperfect information games and other sequential game settings. We then investigate how adapting the reward (by adding a regularization term) of the game can give strong convergence guarantees in monotone games. We continue by showing how this reward adaptation technique can be leveraged to build algorithms that converge exactly to the Nash equilibrium. Finally, we show how these insights can be directly used to build state-of-the-art model-free algorithms for zero-sum two-player Imperfect Information Games (IIG).

From the early days of computing, games have been important testbeds for studying how well machines can do sophisticated decision making. In recent years, machine learning has made dramatic advances with artificial agents reaching superhuman performance in challenge domains like Go, Atari, and some variants of poker. As with their predecessors of chess, checkers, and backgammon, these game domains have driven research by providing sophisticated yet well-defined challenges for artificial intelligence practitioners. We continue this tradition by proposing the game of Hanabi as a new challenge domain with novel problems that arise from its combination of purely cooperative gameplay with two to five players and imperfect information. In particular, we argue that Hanabi elevates reasoning about the beliefs and intentions of other agents to the foreground. We believe developing novel techniques for such theory of mind reasoning will not only be crucial for success in Hanabi, but also in broader collaborative efforts, especially those with human partners. To facilitate future research, we introduce the open-source Hanabi Learning Environment, propose an experimental framework for the research community to evaluate algorithmic advances, and assess the performance of current state-of-the-art techniques.

Learning strategies for imperfect information games from samples of interaction is a challenging problem. A common method for this setting, Monte Carlo Counterfactual Regret Minimization (MCCFR), can have slow long-term convergence rates due to high variance. In this paper, we introduce a variance reduction technique (VR-MCCFR) that applies to any sampling variant of MCCFR. Using this technique, periteration estimated values and updates are reformulated as a function of sampled values and state-action baselines, similar to their use in policy gradient reinforcement learning. The new formulation allows estimates to be bootstrapped from other estimates within the same episode, propagating the benefits of baselines along the sampled trajectory; the estimates remain unbiased even when bootstrapping from other estimates. Finally, we show that given a perfect baseline, the variance of the value estimates can be reduced to zero. Experimental evaluation shows that VR-MCCFR brings an order of magnitude speedup, while the empirical variance decreases by three orders of magnitude. The decreased variance allows for the first time CFR+ to be used with sampling, increasing the speedup to two orders of magnitude.

When observing the actions of others, humans make inferences about why they acted as they did, and what this implies about the world; humans also use the fact that their actions will be interpreted in this manner, allowing them to act informatively and thereby communicate efficiently with others. Although learning algorithms have recently achieved superhuman performance in a number of two-player, zero-sum games, scalable multi-agent reinforcement learning algorithms that can discover effective strategies and conventions in complex, partially observable settings have proven elusive. We present the Bayesian action decoder (BAD), a new multi-agent learning method that uses an approximate Bayesian update to obtain a public belief that conditions on the actions taken by all agents in the environment. BAD introduces a new Markov decision process, the public belief MDP, in which the action space consists of all deterministic partial policies, and exploits the fact that an agent acting only on this public belief state can still learn to use its private information if the action space is augmented to be over all partial policies mapping private information into environment actions. The Bayesian update is closely related to the theory of mind reasoning that humans carry out when observing others’ actions. We first validate BAD on a proof-of-principle two-step matrix game, where it outperforms policy gradient methods; we then evaluate BAD on the challenging, cooperative partial-information card game Hanabi, where, in the two-player setting, it surpasses all previously published learning and hand-coded approaches, establishing a new state of the art.

The CFR+ algorithm for solving imperfect information games is a variant of the popular CFR algorithm, with faster empirical performance on a range of problems. It was introduced with a theoretical upper bound on solution error, but subsequent work showed an error in one step of the proof. We provide updated proofs to recover the original bound.

Evaluating agent performance when outcomes are stochastic and agents use randomized strategies can be challenging when there is limited data available. The variance of sampled outcomes may make the simple approach of Monte Carlo sampling inadequate. This is the case for agents playing heads-up no-limit Texas hold'em poker, whereman-machine competitions typically involve multiple days of consistent play by multiple players, but still can (and sometimes did) result in statistically insignificant conclusions. In this paper, we introduce AIVAT, a low variance, provably unbiased value assessment tool that exploits an arbitrary heuristic estimate of state value, as well as the explicit strategy of a subset of the agents. Unlike existing techniques which reduce the variance from chance events, or only consider game ending actions, AIVAT reduces the variance both from choices by nature and by players with a known strategy. The resulting estimator produces results that significantly outperform previous state of the art techniques. It was able to reduce the standard deviation of a Texas hold'em poker man-machine match by 85\% and consequently requires 44 times fewer games to draw the same statistical conclusion. AIVAT enabled the first statistically significant AI victory against professional poker players in no-limit hold'em.Furthermore, the technique was powerful enough to produce statistically significant results versus individual players, not just an aggregate pool of the players. We also used AIVAT to analyze a short series of AI vs human poker tournaments,producing statistical significant results with as few as 28 matches.

Decomposition, i.e. independently analyzing possible subgames, has proven to be an essential principle for effective decision-making in perfect information games. However, in imperfect information games, decomposition has proven to be problematic. To date, all proposed techniques for decomposition in imperfect information games have abandoned theoretical guarantees. This work presents the first technique for decomposing an imperfect information game into subgames that can be solved independently, while retaining optimality guarantees on the full-game solution. We can use this technique to construct theoretically justified algorithms that make better use of information available at run-time, overcome memory or disk limitations at run-time, or make a time/space trade-off to overcome memory or disk limitations while solving a game. In particular, we present an algorithm for subgame solving which guarantees performance in the whole game, in contrast to existing methods which may have unbounded error. In addition, we present an offline game solving algorithm, CFR-D, which can produce a Nash equilibrium for a game that is larger than available storage.

Extensive-form games are a powerful tool for representing complex multi-agent interactions. Nash equilibrium strategies are commonly used as a solution concept for extensive-form games, but many games are too large for the computation of Nash equilibria to be tractable. In these large games, exploitability has traditionally been used to measure deviation from Nash equilibrium, and thus strategies are aimed to achieve minimal exploitability. However, while exploitability measures a strategy's worst-case performance, it fails to capture how likely that worst-case is to be observed in practice. In fact, empirical evidence has shown that a less exploitable strategy can perform worse than a more exploitable strategy in one-on-one play against a variety of opponents. In this work, we propose a class of response functions that can be used to measure the strength of a strategy. We prove that standard no-regret algorithms can be used to learn optimal strategies for a scenario where the opponent uses one of these response functions. We demonstrate the effectiveness of this technique in Leduc Hold'em against opponents that use the UCT Monte Carlo tree search algorithm.

In large extensive form games with imperfect information, Counterfactual Regret Minimization (CFR) is a popular, iterative algorithm for computing approximate Nash equilibria. While the base algorithm performs a full tree traversal on each iteration, Monte Carlo CFR (MCCFR) reduces the per iteration time cost by traversing just a sampled portion of the tree. On the other hand, MCCFR's sampled values introduce variance, and the effects of this variance were previously unknown. In this paper, we generalize MCCFR by considering any generic estimator of the sought values. We show that any choice of an estimator can be used to probabilistically minimize regret, provided the estimator is bounded and unbiased. In addition, we relate the variance of the estimator to the convergence rate of an algorithm that calculates regret directly from the estimator. We demonstrate the application of our analysis by defining a new bounded, unbiased estimator with empirically lower variance than MCCFR estimates. Finally, we use this estimator in a new sampling algorithm to compute approximate equilibria in Goofspiel, Bluff, and Texas hold'em poker. Under each of our selected sampling schemes, our new algorithm converges faster than MCCFR.

Extensive-form games are a powerful model for representing interactions between agents. Nash equilibrium strategies are a common solution concept for extensive-form games and, in two-player zero-sum games, there are efficient algorithms for calculating such strategies. In large games, this computation may require too much memory and time to be tractable. A standard approach in such cases is to apply a lossy state-space abstraction technique to produce a smaller abstract game that can be tractably solved, while hoping that the resulting abstract game equilibrium is close to an equilibrium strategy in the unabstracted game. Recent work has shown that this assumption is unreliable, and an arbitrary Nash equilibrium in the abstract game is unlikely to be even near the least suboptimal strategy that can be represented in that space. In this work, we present for the first time an algorithm which efficiently finds optimal abstract strategies --- strategies with minimal exploitability in the unabstracted game. We use this technique to find the least exploitable strategy ever reported for two-player limit Texas hold'em.

Counterfactual Regret Minimization (CFR) is a popular, iterative algorithm for computing strategies in extensive-form games. The Monte Carlo CFR (MCCFR) variants reduce the per iteration time cost of CFR by traversing a sampled portion of the tree. The previous most effective instances of MCCFR can still be very slow in games with many player actions since they sample every action for a given player. In this paper, we present a new MCCFR algorithm, Average Strategy Sampling (AS), that samples a subset of the player's actions according to the player's average strategy. Our new algorithm is inspired by a new, tighter bound on the number of iterations required by CFR to converge to a given solution quality. In addition, we prove a similar, tighter bound for AS and other popular MCCFR variants. Finally, we validate our work by demonstrating that AS converges faster than previous MCCFR algorithms in both no-limit poker and Bluff.

In this paper we show that the move pruning method we presented at SoCS last year sometimes prunes all the least-cost paths from one state to another. We present two examples exhibiting this erroneous behaviour--a simple, artificial example and a slightly more complex example that arose in last year's experiments. We then formally prove that a simple modification to our move pruning method makes it "safe," i.e., it will never prune all the least-cost paths between a pair of states. Finally, we present the results of rerunning last year's experiments with this provably safe version of move pruning and show that last year's main conclusions still hold, namely: (1) in domains where there are short redundant sequences move pruning produces substantial speedups in depth-first search, and (2) in domains where the only short redundant sequences are 2-cycles, move pruning is faster than parent pruning by a factor of two or more.

We present three new ideas for grid-based path-planning algorithms that improve the search speed and quality of the paths found. First, we introduce a new type of database, the Local Distance Database (LDDB), that contains distances between boundary points of a local neighborhood. Second, an LDDB based algorithm is introduced, called Block A*, that calculates the optimal path between start and goal locations given the local distances stored in the LDDB. Third, our experimental results for any-angle path planning in a wide variety of test domains, including real game maps, show that Block A* is faster than both A* and the previously best grid-based any-angle search algorithm, Theta*.

Move pruning is a low-overhead technique for reducing the size of a depth first search tree. The existing algorithm for automatically discovering move pruning information is restricted to games where all moves can be applied to every state. This paper demonstrates an algorithm which handles a general class of single player games. It gives experimental results for our technique, demonstrating both the applicability to a range of games, and the reduction in search tree size. We also provide some conditions under which move pruning is safe, and when it may interfere with other search reduction techniques.

We present three new ideas for grid-based path-planning algorithms that improve the search speed and quality of the paths found. First, we introduce a new type of database, the Local Distance Database (LDDB), that contains distances between boundary points of a local neighborhood. Second, an LDDB-based algorithm is introduced, called Block A*, that calculates the optimal path between start and goal locations given the local distances stored in the LDDB. Third, our experimental results for any-angle path planning in a wide varietyof test domains, including real game maps, show that Block A* is faster than both A* and the previously best grid-based any-angle search algorithm, Theta*.

An important part of the creation of a housing subdivision is the design and layout of sewers underneath the road. This is a challenging cost optimization problem in a continuous threedimensional space. In this paper, heuristic-search-based techniques are proposed for tackling this problem. The result is new algorithms that can quickly find near optimal solutions that offer important reductions in the cost of design and construction.

Korf, Reid, and Edelkamp introduced a formula to predict the number of nodes IDA* will expand on a single iteration for a given consistent heuristic, and experimentally demonstrated that it could make very accurate predictions. In this paper we show that, in addition to requiring the heuristic to be consistent, their formula's predictions are accurate only at levels of the brute-force search tree where the heuristic values obey the unconditional distribution that they defined and then used in their formula. We then propose a new formula that works well without these requirements, i.e., it can make accurate predictions of IDA*'s performance for inconsistent heuristics and if the heuristic values in any level do not obey the unconditional distribution. In order to achieve this we introduce the conditional distribution of heuristic values which is a generalization of their unconditional heuristic distribution. We also provide extensions of our formula that handle individual start states and the augmentation of IDA* with bidirectional pathmax (BPMX), a technique for propagating heuristic values when inconsistent heuristics are used. Experimental results demonstrate the accuracy of our new method and all its variations.

Typically agent evaluation is done through Monte Carlo estimation. However, stochastic agent decisions and stochastic outcomes can make this approach inefficient, requiring many samples for an accurate estimate. We present a new technique that can be used to simultaneously evaluate many strategies while playing a single strategy in the context of an extensive game. This technique is based on importance sampling, but utilizes two new mechanisms for significantly reducing variance in the estimates. We demonstrate its effectiveness in the domain of poker, where stochasticity makes traditional evaluation problematic.