Skip to content

README.md

Latest commit

 

History

History
29 lines (21 loc) · 1.35 KB

README.md

File metadata and controls

29 lines (21 loc) · 1.35 KB

Claw Machine Experiment Data Modelling

Modelling claw machine experiment data using Optimal Bayesian decision-making agent. Confidence is given by the entropy of the Gibb's posterior.

  • Each category is a univariate Gaussian N(μ, σ).
  • On a trial the agent is offered two categories and must choose which one to sample from.
  • An outcome is a success if the draw is below a fixed threshold (T), otherwise a failure.
  • Rewards can differ between categories and are applied on success only.
  • The agent maximises expected exponential utility

Bayesian decision-making process

In main section of decision_maker.py:

- Adjust parameters of the ball distributions via "cats" dictionary
- Adjust rewards of the ball distributions via "rews" dictionary
- Provide pairs to evaluate via list of tuples called "pairs"
- When initialising the agent we need:
    - Threshold, T
    - Inverse temperature parameter, β
    - Decision-rule flag: if soft = False, then we do a MAP decision rule, otherwise soft sampling of the catgeory

  • soft=True makes the agent sample from the posterior over actions.
    Set soft=False for the deterministic MAP choice. Adjust β parameter to adjust risk sensitivity of agent

  • confidence is the negative entropy of that posterior (larger = more confident).