TIL - shivamvats/notes GitHub Wiki

  • Huber loss. Loss function more robust to outliers used in neural networks, e.g. DQN.
  • Save my figures as an svg as I can edit and improve the figure later in inkscape.
  • Mixture of Bernoulli Distributions: Details
  • Isotonic Regression: fit a monotonically increasing line to data.
  • Riemann-Stieltjes Integration in Probability theory: The Riemann-Stieltjes integral generalizes the concept of the Riemann integral and allows for integration with respect to a cumulative distribution function that isn't continuous. Details
  • Borel Paradox: The concept of conditional probability distribution for continuous random variables is not very intuitive. This paradox shows that conditional density functions need not be invariant under coordinate transformations. Details
  • Zero-sum Stochastic Game: is a model general model than the MDP.
  • If you want to get reasonably good at something, deconstruct it and spend 20 hours on it distraction free.
  • Being distracted is just a symptom. The root cause is over-stimulation of the brain. To enjoy the little joys of life and self-improvement, remove sources of stimulation from your life.
  • Kolmogorov-Smirnov (K-S) Test: is a test of the equality of two 1d continuous or discrete probability distributions. The one sample K-S test is used to compare a sample with a reference probability distribution. The two sample K-S test is used to test the equality of two samples.
  • Epistemic Planning: is the enrichment of planning with epistemic notions, such as, knowledge and belief, formalized using an epistemic logic. See details
  • Null Space Projection: Useful when running a set of controllers together to ensure that they don't produce opposing wrenches.
  • vuja de: something familiar viewed with a fresh perspective (opposite of deja vu)
  • Levin Tree Search: It uses a learned action policy to guide search instead of a heuristic function and guarantees an upper bound on search steps based on the quality of the policy.
  • Aleatory Variability refers to the natural randomness in a variable while Epistemic Uncertainty refers to uncertainty in the model of a process due to limited data.
  • Multimodal-Optimization: The problem of finding multiple/all high quality locally optimal solutions. See details.
  • Totally Unimodular Matrix: If the constraints of a linear program can be represented by a totally unimodular (TU) matrix, then the LP is integral. In other words, we convert an Integer Program (NP-complete) to an LP (P) if its constraint matrix is TU. See details
  • Bag of Features: Inspired by the Bag of words model used in NLP. Details.