Decentralized partially-observable Markov decision processes (Dec-POMDPs) are a powerful tool for modeling multi-agent planning and decision-making under uncertainty. Prevalent Dec-POMDP solution techniques require centralized computation given full knowledge of the underlying model. Multi-agent reinforcement learning (MARL) based approaches have been recently proposed for distributed solution of Dec-POMDPs without full prior knowledge of the model, but these methods assume that conditions during learning and policy execution are identical. In some practical scenarios this may not be the case. We propose a novel MARL approach in which agents are allowed to rehearse with information that will not be available during policy execution. The key is for the agents to learn policies that do not explicitly rely on these rehearsal features. We also establish a weak convergence result for our algorithm, RLaR, demonstrating that RLaR converges in probability when certain conditions are met. We show experimentally that incorporating rehearsal features can enhance the learning rate compared to non-rehearsal-based learners, and demonstrate fast, (near) optimal performance on many existing benchmark Dec-POMDP problems. We also compare RLaR against an existing approximate Dec-POMDP solver which, like RLaR, does not assume a priori knowledge of the model. While RLaR's policy representation is not as scalable, we show that RLaR produces higher quality policies for most problems and horizons studied.

No abstract

No abstract

We present new Multiagent learning (MAL) algorithms with the general philosophy of policy convergence against some classes of opponents but otherwise ensuring high payoffs. We consider a 3-class breakdown of opponent types: (eventually) stationary, self-play and "other" (see Definition 4) agents. We start with ReDVaLeR that can satisfy policy convergence against the first two types and no-regret against the third, but it needs to know the type of the opponents. This serves as a baseline to delineate the difficulty of achieving these goals. We show that a simple modification on ReDVaLeR yields a new algorithm, RV σ (t) , that achieves no-regret payoffs in all games, and convergence to Nash equilibria in self-play (and to best response against eventually stationary opponents -a corollary of no-regret) simultaneously, without knowing the opponent types, but in a smaller class of games than ReDVaLeR . RV σ (t) effectively ensures the performance of a learner during the process of learning, as opposed to the performance of a learned behavior. We show that the expression for regret of RV σ (t) can have a slightly better form than those of other comparable algorithms like GIGA and GIGA-WoLF though, contrastingly, our analysis is in continuous time. Moreover, experiments show that RV σ (t) can converge to an equilibrium in some cases where GIGA, GIGA-WoLF would fail, and to better equilibria where GIGA, GIGA-WoLF converge to undesirable equilibria (coordination games). This important class of coordination games also highlights the key desirability of policy convergence as a criterion for MAL in self-play instead of high average payoffs. To our knowledge, this is also the first successful (guaranteed) attempt at policy convergence of a no-regret algorithm in the Shapley game.

When an agent learns in a multi-agent environment, the payoff it receives is dependent on the behaviour of the other agents. If the other agents are also learning, its reward distribution becomes non-stationary. This makes learning in multi-agent systems more difficult than single-agent learning. Prior attempts at value-function based learning in such domains have used off-policy Q-learning that do not scale well as the cornerstone, with restricted success. This paper studies on-policy modifications of such algorithms, with the promise of scalability and efficiency. In particular, it is proven that these hybrid techniques are guaranteed to converge to their desired fixed points under some restrictions. It is also shown, experimentally, that the new techniques can learn (from self-play) better policies than the previous algorithms (also in self-play) during some phases of the exploration.