COSMOS 2023

Computational summer school on modeling social and collective behavior - Konstanz (DE) July 4th - 7th

Group projects

The general concept

• Get hands-on experience with the concepts introduced during talks and in tutorials
• ~5 hours to work on your project (Thursday & Friday after lunch)
• Present results/how things went (Friday afternoon)
• Groups of 5-6
• Beginner friendly projects are marked 🫑, advanced are 🌶️

P1. Wisdom or madness of crowds?

• It has been shown that social influence can be both beneficial (‘wisdom of crowds’) or maladaptive (‘herding’)
• But when does which effect occur?

Theoretical background

Based on Toyokawa, Whalen, & Laland (2019)

vs.

🫑

P1. Wisdom or madness of crowds?

The task

The data & existing code

• Data from participants performing the task
• Chosen slot machine, reward, social frequency of the machines, optimal choice, and group size in the data
• R code implementing a social learning agent that uses a mixture of individual (Q-learning) and social (frequency-based) influence to make choices

🫑

P1. Wisdom or madness of crowds?

Possible research questions/projects

• Try to predict human choice behaviour
• Try out the conformity-biased Q-learning from the paper and compare it to alternative models from literature (or your own!)
• Is social learning beneficial or harmful in this case?

The model - choice probability for option m by individual i in trial t

σ_i,t - social learning weight (∈[0;1]); θ_i - conformity exponent (∈ [−∞,+∞], α_i - learning rate (∈[0;1]), r_i,t(m) - monetary reward from machine m in trial t, 1(m, m_i,t) - indicator function determining if option m was chosen in trial t

🫑

P2. Social learning in correlated environments

?

☠️

: (

vs.

• Much of social learning research has focused on learning from others when reward functions are identical
• However, in real life we often learn from others with different skills, preferences, or goals
• How do humans integrate social information when it is only positively correlated, and not identical to, their own reality?

Theoretical background

Based on Witt, Toyokawa, Lala, Gaissmaier, & Wu, 2023

🌶️

P2. Social learning in correlated environments

The task

The data & existing code

• Data from 3 groups of participants performing the task
• Chosen options, rewards, search distance from previous trial, distance from and previous rewards of other players
• Python code implementing different variants of GP-UCB agents (which have different methods of integrating social information)

🌶️

P2. Social learning in correlated environments

Possible research questions/projects

• Try to predict human search behaviour
• Find the best fitting model out of the existing ones
• Come up with your own
• Find a model that weights different participants’ info differently (did a participant trust some fellow players more than others?)
• Do humans use social information at all? Do they rely on it more or less than individual information?

The models

• Asocial Learning (AS) - only uses individual info
• Decision Biasing (DB) - mixes a frequency-based social and individual policy
• Value Shaping (VS) - prediction error updates to value function based on social information
• Social Generalization (SG) - generalizes social as well as individual information, but treats social information as noisier

🌶️

P3. Learning other people’s values based on the decisions they make

• The ability to figure out what other people like or value is critical for social reasoning
• Inferring people’s preferences requires paying attention to the context of people’s choices (including available alternatives and the costs involved)
• This project teaches you the most basic building block of Bayesian Theory of Mind models, including implementing a basic generative model, and performing Bayesian inference over it.

Theoretical background

🫑

Based on Jern, Lucas, & Kemp (2017)

P3. Learning other people’s values based on the decisions they make

The task

Implementation context

• Choice context (available options and their costs), and chosen option
• No code available. You will need to know how to write functions, write loops, use basic sampling functions (e.g., sample random value), and use data structures (matrices).

🫑

P3. Learning other people’s values based on the decisions they make

Possible research questions/projects

• Simulate data from an artificial agent making sequential choices, and have an observer infer the agent’s preferences
• Use the inferred preferences to predict future behavior
• Compare the models!
• Extend the models to also infer how rational the agent is being

The models

🫑

• Simple Associations - choices directly reveal preferences
• Sampling Sensitivity - consider the space of other available choices
• Cost Sensitivity - consider the relative differences in costs

P4. Tracking competence and preference inference from spatial navigation

• When you see people navigating in space, how do you figure out what they’re trying to accomplish and what they are good or bad at?
• How can you then use these inferences to predict someone’s next actions, decide if they need help, or determine if their intentions are good or bad?

Theoretical background

Based on Jara-Ettinger, Schulz & Tenenbaum (2021)

🌶️

P4. Tracking competence and preference inference from spatial navigation

The task

Implementation context

• You will have the data from Jara-Ettinger et al., (2021), which has participant inferences from an agent moving around in a 2D world. Participants inferred past preferences and future actions.
• A codebase is available at http://github.com/julianje/Bishop, implemented in Python. You can either use this codebase, or implement your own Markov Decision Process (MDP).

🌶️

Based on Jara-Ettinger, Schulz & Tenenbaum (2020)

P4. Tracking competence and preference inference from spatial navigation

Possible research questions/projects

• Simulate data from an artificial agent navigating an environment, and have an observer infer its competence and preferences
• Compare your model predictions to participant judgments
• Come up with ways to extend the model to handle other types of social situations!

🌶️

Based on Jara-Ettinger, Schulz & Tenenbaum (2021)

P5. Learning to communicate about shared procedural abstractions

Theoretical background

​

Based on McCarthy*, Hawkins*, Wang, Holdaway, & Fan (2021)

🫑/🌶️

• Successful collaboration on a project poses multiple challenges, including: (1) how to break complex problems down into simpler ones and (2) how to communicate clearly about what needs to be done.
• How do agents figure out how to do these things together?

P5. Learning to communicate about shared procedural abstractions

The task

Based on McCarthy*, Hawkins*, Wang, Holdaway, & Fan (2021)

🫑/🌶️

P5. Learning to communicate about shared procedural abstractions

Guided reproducibility study 🫑

Based on McCarthy*, Hawkins*, Wang, Holdaway, & Fan (2021)

🫑/🌶️

• Dataset: Language from architect; actions taken by builder
• Notebook: Scaffolded walkthrough to reproduce main results reported in 2021 CogSci paper

Possible extensions (beyond COSMOS?) 🌶️

• Scaling up to wider variety of programs
• Scaling up to more natural language

P6. Modelling algorithm-mediated social learning

Theoretical background

​

Inspired by Acerbi (2019), Brady et al. (2023)

🫑

• Transmission biases / social learning strategies are a central concept in cultural evolution
• When cultural transmission happens online (with other humans on social media, or with AI agents) they interact with algorithms that filter the information we receive.

P6. Modelling algorithm-mediated social learning

Research questions

​

🫑

• Extending transmission biases / social learning strategies models including forms of algorithmic mediation, and studying population-level consequences.
• Possible data from social media?

P6. Modelling algorithm-mediated social learning

Existing code

​

🫑

• We can expand codes from Acerbi, Mesoudi, Smolla (2022), available online at: https://acerbialberto.com/IBM-cultevo/

P7. Interplay between direct learning and social learning under uncertainty

Theoretical background

​

based on Zhang & Gläscher (2020)

• Humans learn from themselves (direct learning) and also from other people (social learning)
• Less is known about how these two types of learning interacts when multiple individuals (e.g., N=5) are engaged in the same learning environment
• Social information pre-/post- reward feedback should be dissociated

🌶️

P7. Interplay between direct learning and social learning under uncertainty

Data & Model

​

🌶️

P7. Interplay between direct learning and social learning under uncertainty

Research questions

🌶️

Existing code: https://github.com/lei-zhang/SIT

• Understand meanings of model parameters
• Predict human choices (aka, posterior predictive checks)
• Try to simulate different “social agents” for a hypothetical experiment with “pretended” social partners

P8. From pattern to process

based on Kandler, Fogarty & Karsdorp (2023)

Theoretical background:

​

• One focus of cultural evolutionary research is understanding how individuals learn socially

​

• Due to the fact that many archaeological and anthropological data sets record the occurrence/usage of different variants of a cultural trait in the population, this often involves solving an inverse problem.

🫑

P8. From pattern to process

The task:

​

There will be 5 data sets describing the occurrence of different variants of a cultural trait in a sample taken from a population of 100 individuals.

Which data sets have been generated through a process of unbiased cultural transmission, i.e. the process where the probability of choosing any cultural variant to adopt is dictated only by its relative frequency in the pool of potential role models?

🫑

P8. From pattern to process

Possible research questions:

​

• How do you apply cultural evolutionary theory to such data (code implementing the process of unbiased cultural transmission will be available)?

​

• What biases or constraints can you think of that may interfere with such inferential studies?

​

• How would you include those biases or constraints in the inferential procedure? Does it change your results?

🫑

P9. Quantifying leadership in animal groups

Theoretical background:

​

• Animal behaviours exhibit complex temporal dynamics, suggesting that there are multiple timescales at which they should be studied

​

• However, research on animal movement has tended to focus on relatively restricted temporal scales, typically ones most accessible to observation or data collection.

​

based on Sridhar et al. (2023)

🫑

P9. Quantifying leadership in animal groups

​

The task:

​

There will be 2 datasets of animal groups—schools of golden shiner fish and flocks of pigeons—that provide relational position and kinematics between dyads, and also a binary variable indicating if the pair is engaged in a leader-follower interaction, or not.

You can explore the effect of different pairwise predictors on leadership / social influence within the group

​

🫑

P9. Quantifying leadership in animal groups

​

Possible research questions:

​

• What factors affect leadership within animal groups?
• What insights are gained when these data are analysed at different timescales?
• What is the distribution of social influence within the group?
• What species specific differences are captured by these analyses?

​

🫑

Time to choose!

P1: Wisdom or madness of crowds 🫑

P2: Social learning in correlated environments 🌶️

P3: Learning people’s values based on the decisions they make 🫑

P4: Tracking competence and preference inference from spatial navigation 🌶️

P5: Learning to communicate about shared procedural abstractions 🫑 /🌶️

P6: Modelling algorithm-mediated social learning 🫑

P7: Interplay between direct learning and social learning under uncertainty 🌶️

P8: From pattern to process 🫑

P9: Quantifying leadership in animal groups

Have fun!