Originally posted by **hyunjimoon** February 22, 2024
using ten standards on good bayesian model, we can compare learning algorithms of each startup based on business model
1. Causal Consistency: Emphasizes the need for operations in distributed systems to maintain a causally consistent order.
2. Parameter Recoverability: Discusses methods for accurately retrieving model parameters from data.
3. Predictive Performance: Evaluates models' accuracy in forecasting future outcomes.
4. Fairness: Addresses the need for equitable outcomes in models and algorithms.
5. Structural Faithfulness: Explores the alignment of models with the true data structure.
6. Parsimony: Advocates for simplicity in models for effective explanation and prediction.
7. Interpretability: Highlights the importance of making models understandable to humans.
8. Convergence: Reviews how and when algorithms stabilize to a solution.
9. Estimation Speed: Focuses on increasing the efficiency of estimation processes.
10. Robustness: Examines models' reliability under various conditions.
examples:
- just as gibbs sampler perform better in low causal density, slower ABC iteration (AAABBBCCC) is better in industry with high atomness due to its high "shifting gear costs". For high causal density industry, hmc sampler (ABCABCABC) is better for high causal density , due to low "shifting gear cost" which is relevant to 8
- horseshoe prior is adding sparsity to improve 7 and 9
Discussed in https://github.com/Data4DM/BayesSD/discussions/191
todo: complete summarizing
2.1 iai1,2,3(solo vs duo)
2.2 agent and environment's interaction using sampling (MCMC) algorithm
2.3 alignment, evaluation, experiment design dynamics