Introducing
Your new presentation assistant.
Refine, enhance, and tailor your content, source relevant images, and edit visuals quicker than ever before.
Trending searches
Reinforcement Communication Learning in Different Social Net
9
Research motivation
Reinforcement Communication Learning in Different Social Network Structures
How did human languages evolve?
Goal: building a rich and open-ended model of language evolution
Using AI for Modeling Multi-Agent Communication Learning
Results
- Instability
- Developers often centralize the optimization to help models to converge on stable communication
We know: human groups quickly learn new communicative conventions in "decentralized" experimental settings
Solution: guiding model development by empirical evidence on individual- and population-level factors that drive decentralized learning in human groups
Marina Dubova, Arseny Moskvichev & Robert Goldstone
mdubova@iu.edu
@dubova_marina
Human Learning in Different Social Network Structures
Social network structure shapes the communication systems developed in the groups of RL agents:
- Agent degrees affect how consistently they use the communication channel
- Global social connections force the agents to converge on shared and symmetric communication systems
Reward scores are not enough! A comprehensive set of metrics is required to reliably interpret the results of multi-agent communication learning.
Social network structure shapes decentralized learning in human groups:
- local connectivity supports independent exploration
- global connectivity helps to converge on a shared solution
Naming game experiment (Centola & Baronchelli, 2015):
- social network with many local connections ("ring" structure) or in a randomly-connected network -> many local conventions
- fully-connected network ("clique") -> global consensus
Experiment 3: proportion of global connections
- Local social connections led to the emergence of local and asymmetric communication systems (“dialects”)
- Adding more global connections forced agents to find global consensus
Coordination Game
Experiment 2: average degree
Average degree of a social network is a central factor affecting how deterministically agents use the communication channel
Social Network Types
random
ring
small-world
clique
(fully-connected)
Experiment 1
Metrics
- Random and fully-connected social networks -> symmetric and homogeneous communication systems
- Small-world and ring social networks -> asymmetric and local communication systems
- Agents in fully-connected social network used communication channel less deterministically
- Agents in fully-connected and random networks produced less diversified actions, but their action and signaling distributions were much more coordinated than in the ring-shaped and small-world network structures