Cooperative AI is a subfield of artificial intelligence that aims to improve the ability of AI systems to engender cooperation between humans, machines, and institutions. In this work, I'm specifically interested in exploring the dynamics between agents representing social institutions like organizations, companies, or even areas of government and policy decisions within the domains of Cooperative AI. I want to understand how heterogeneous agents, that posses largely different skills and characteristics, like different economic power, political influence, different levels of organization, and different comprehension of their environment, can impact the way they these agents with such asymmetric capabilities interact and cooperate with one another.
keywords: Multi-agent reinforcement learning; Evolutionary game theory; Large Language Models
In the past few months, I have been conducting research to understand the patterns of interactions between such heterogeneous agents within general sum game contexts in Multi-Agent Reinforcement Learning environments [1]. This research builds upon previous explorations [2,3,4] that establish a link between the findings of evolutionary biologists like those in Nowak's seminal work on the "Five Rules for the Evolution of Cooperation" [5,6], to works found in game theory, such as the algorithms demonstrated in Axelrod's "The Evolution of Cooperation" [7], to Reinforcement Learning scenarios, in which agents pursue what we believe to be similar solutions to social dilemmas than the ones observed in biology and sociology.
The perspective brought by the use of this new framework might bring a revolution to the area of the study of cooperation. While previous studies were limited by the researchers' understanding and confined to finite representations of discrete states and actions within deterministic Markov Processes dynamics, Reinforcement Learning opens the doors for a much wider exploration. It allows researchers to experiment with Deep Neural Network layers that enable a continuous exploration and action space, permitting far more complex modeling of highly dynamical systems. These could include interactions between companies, economical models of a city, or decisions a state can make.
Despite the inherent uncertainties, the alignment of these models with established principles from other disciplines provides a compelling reason to explore the potential insights they could offer into the intricate decision-making processes and emergent behaviors that shape cooperation among diverse entities with varying capabilities and constraints.
Modeling Agents with Deep Reinforcement Learning
This research aims to leverage deep reinforcement learning (DRL) to model the intricate dynamics that govern interactions between heterogeneous agents. Equipped with actor-critic architectures, these agents can adapt their strategies in response to the evolving behaviors of other agents within the environment. Crucially, they possess the capability to consider both immediate rewards and future states, a vital trait for fostering sustainable cooperation. Moreover, the versatility of DRL allows for the simulation of highly complex, multidimensional scenarios that approximate real-world dynamics by incorporating a myriad of variables. In these rich virtual environments, agents must grapple with decision-making under uncertainty and incomplete information, mirroring the challenges faced in real scenarios.
The problem of cooperation is most pronounced in social dilemmas, where individual incentives and collective welfare are inherently misaligned. Individuals may be enticed to exploit others for personal gain or act defensively out of fear of being exploited themselves. However, collective welfare is maximized when all parties choose to cooperate [8]. To investigate the preconditions that foster cooperative behavior, I intend to employ a rich, multi-dimensional environment capable of simulating general-sum game interactions akin to the Prisoner's Dilemma. This classical model serves as an ideal framework for delineating the fundamental tensions between individual incentives and collective welfare.
While it may be challenging to establish direct connections between reinforcement learning systems and typical game-theoretical and evolutionary biology environments, it is certainly possible [9]. We will maintain the guiding principle that a cooperator is an entity that pays a cost, c, for another individual to receive a benefit, b, while a defector incurs no cost and does not provide benefits [2]. Cost and benefit are measured in terms of the agents' utility [10].
While the application of DRL in simulating economic interactions has its upsides, the complexity of economic systems clearly also poses challenges. Accurately modeling the nuanced behaviors of economic agents and the unpredictable nature of global economic interactions is an unfeasible task. I aim to develop a model.
This research is not intended to replace the work of economists, historians, or sociologists. Instead, the goal is to augment their toolkit with AI-driven analyses that can offer alternative perspectives and richer data-driven insights.
Through attending the Complexity Global School, I aim to deepen my understanding of the economic realities faced by various countries and institutions and gain a more complete picture on what are the important variables to represent in this research environment.