Prosocial learning agents solve generalized Stag Hunts better than selfish ones

By Alex Peysakhovich, Adam Lerer. 2017

127 citations recorded by the source corpus.

Real world interactions are full of coordination problems [2, 3, 8, 14, 15] and thus constructing agents that can solve them is an important problem for artificial intelligence research. One of the simplest, most heavily studied coordination problems is the matrixform, two-player Stag Hunt. In the Stag Hunt, each player makes a choice between a risky action (hunt the stag) and a safe action (forage for mushrooms). Foraging for mushrooms always yields a safe payoff while hunting yields a high payoff if the other player also hunts but a very low payoff if one shows up to hunt alone. This game has two important Nash equilibria: either both players show up to hunt (this is called the payoff dominant equilibrium) or both players stay home and forage (this is called the risk-dominant equilibrium [7]).

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Citations in the corpus

Discovering Diverse Multi-Agent Strategic Behavior via Reward Randomization (2021, methodology): “We consider two games adapted from Peysakhovich & Lerer (2018b), Monster-Hunt (Fig.”
Relational Forward Models for Multi-Agent Learning (2018, positive): “We considered three multi-agent environments for our study: Cooperative Navigation (Lowe et al., 2017), Coin Game (Raileanu et al., 2018) and Stag Hunt (Peysakhovich & Lerer, 2017b).”
Emergent Communication through Negotiation (2018, background): “These different reward schemes can be seen as speciﬁc instantiations of a more general reward formula of the form R = αR A + βR B (Peysakhovich & Lerer, 2017).”
Social Diversity and Social Preferences in Mixed-Motive Reinforcement Learning (2020, background): “These resemble previous approaches using selfishness [42], inequity aversion [24, 56], and strong reward sharing [44, 52], respectively.”
Influence-Based Multi-Agent Exploration (2019, methodology): “Island: This task is a modiﬁed version of the classic Stag Hunt game (Peysakhovich & Lerer, 2018) where two agents roam a 10 × 10 island populated with 9 treasures and a random walking beast for 300 environment steps.”
Consequentialist conditional cooperation in social dilemmas with imperfect information (2017, background): “…(Foerster et al. 2017a), multi-agent coordination problems (Lowe et al. 2017; Foerster et al. 2017b; Riedmiller et al. 2009), (Tampuu et al. 2017; Peysakhovich and Lerer 2017), and the emergence of language (Lazaridou, Peysakhovich, and Baroni 2017; Das et al. 2017; Evtimova et al. 2017;…”
Inequity aversion improves cooperation in intertemporal social dilemmas (2018, background): “planning [20, 21, 22, 23] to generate cooperation in this setting.”
AllenAct: A Framework for Embodied AI Research (2020, methodology): “As in [57], the Stag moves towards to the closet agent at each step.”