Use Bayesian Inference to measure uncertainty and Lipschitz to cap consequences of that uncertainty
The issue with what I see as Bayesian Inference to mimic general intelligence is that it doesn't have a means to focus on an issue in a dynamic or active way. I discovered Lipschitz robustness which I believe solves this issue. Real-world environments are messy. Sensors are noisy, objects move unpredictably, and one bad control command can mean a collision or instability. To build robots that can adapt to any domain , we need both: Bayesian inference — to predict what is likely to happen. Lipschitz robustness — to guarantee what is allowed to happen. This post shows how combining these two ideas gives us robots that are both adaptive and provably safe . Lipschitz Robustness: Motion with Certified Safety Lipschitz continuity ensures that small changes in the input cannot cause huge jumps in the output: ‖f(x') − f(x)‖ ≤ L · ‖x' − x‖ In control terms, it lets us say: If my state estimate or sensor reading is off by a little, then the cont...