Abstract:
Humans and machines increasingly adapt together — over shared resources, over each other, over society itself. Whether these human-machine ecologies thrive or unravel comes down to one question: can many actors, each learning from their own experience and pursuing their own interests, still reach a good joint outcome? We can rarely say which way it will go. Complex systems science captures the emergent behavior of such collectives but tends to simplify the individual, whereas models based on multi-agent reinforcement learning capture rich individual learning but are noisy, costly, and hard to interpret. This talk presents collective learning dynamics, an approach that bridges the two. Instead of running the learning algorithm thousands of times and averaging, we derive the deterministic flow it follows on average and study it using the standard tools of dynamical systems (fixed points, basins of attraction, bifurcations, timescales). The payoff is twofold: results hold across whole families of algorithms rather than one specific update rule, and phenomena invisible in any single run (why a collective locks onto a bad outcome, where the boundary between basins sits) can be read straight off the flow. I will make it concrete on one problem: how do natural tipping dynamics affect sustainable collective action? I will show the cooperation-promoting potential of being embedded in the same, responsive environment, including the conditions that enable cooperation. It's a small, fully worked instance of the question we opened with: which conditions make the good regime the one a human-machine ecology actually reaches.
Bio:
Wolfram Barfuss explores the question of how cooperative behavior arises in which people join forces and act collectively to preserve environmental livelihoods. To do this, he and his team develop mathematical models of collective learning, drawing on ideas from various research areas, including complex adaptive systems, cognitive science, artificial intelligence (particularly multi-agent learning), and social ecology. The goal is to create a new human-environment modeling tool that will be made freely available and continually expanded, so that it can be utilized by researchers from different disciplines.
He is a sustainability system scientist with a doctoral degree in theoretical physics from the Potsdam Institute for Climate Impact Research and the Humboldt University Berlin (2019). Before joining ZEF in 2023, he was a postdoctoral scientist at the Tübingen AI Center at the University of Tübingen (2021-2023), the School of Mathematics at the University of Leeds (2020-2021), and the Max Planck Insitute for Mathematics in the Sciences in Leipzig (2019-2020).
Barfuss heads the Argelander chair of the University of Bonn's Transdisciplinary Research Area (TRA) Sustainable Futures. The Argelander professorships are high-profile tenure-track assistant professor positions newly created within the scope of the German Excellence Initiative. The research focus is Integrated System Modeling for Sustainability Transitions. The BarfussLab is a young, strongly evolving, interdisciplinary team crossing the boundaries of various research fields to identify critical leverage points to tackle collective action challenges for sustainability. How? Through collective learning for a sustainable future.
Summary:
Why: Collection-action strategies
Goal: ecologically safe, socially just and aligned space for society
Need fast large-scale collective action
Scientists report danger and propose solutions
But this is not sufficient to generate collective action
The challenge of collective action / cooperation
Inherent conflict of interest between individuals and the collective
e.g. prisoner’s dilemma: individual benefits from doing anti-social action
Often need central authority to force cooperation
Complex cooperation dynamics:
Poor/no cooperation
fear&greed games
greed&no fear games
fear&no fear games
Full cooperation: No fear & no greed games
Challenge: can many selfish, unique actors cooperate to reach the good life?
Modeling approaches:
No existing model family is appropriate out-of-the-box
Dynamic-systems: no way to capture agency
Equilibrium: no notion of time, assumes rational agent
Agent-based: extremely complex, hard to capture driving phenomena in practice
Building bridges:
Complex Systems Science
Provides coarse insights but can be simplistic
Models interaction networks and how agents will act in different group scenarios
Multi-agent Reinforcement Learning (MARL)
Rich treatment of agent reasoning, but complex and challenging to interpret
Capture learning and iterative decision-making
Can learn to cooperate
Approach for synthesis:
Study low dimensional environments
Reduce stochasticity
Reduce dynamics state space
Reduce feedback asymmetry
Basic framework
Agents take actions and get rewards from environment
Learn how agents affect environment and rewards after many observations
Learn to maximize reward
How: How do natural tipping dynamics affect collective action
Current view: cooperation requires social norms of reciprocity
But having a common environment also drives cooperation via common drivers
Need to isolate ecological interactions in a cooperative game
Game with agents that choose conservation vs intensification
Affects environment, causing degradation or recovery
Depending on how the world is parameterized this world can lead to different categories of social games
Can track the trajectories of agents through the world to understand the structure and timing of tipping dynamics
Although there are no social norms agents can learn to cooperate
Prior experiences cause changes to the parameters of the game
E.g. agents can change how much they care about the future based on past experiences
What: the environment is not a passive backdrop; it is an essential catalyst for forming cooperation