Sequential estimation

Content sourced from Wikipedia, licensed under CC BY-SA 3.0.

Sequential estimation is about estimating things as data comes in, not using a fixed amount of samples. You collect data step by step and stop when a predefined rule says you have enough evidence to be confident about the result. The idea is rooted in Bayesian thinking: you treat the unknowns as evolving over time and update your beliefs as new measurements arrive.

A Markov process models how the hidden state changes over time, and each new measurement provides information about that state. By combining the observed data with the model, you gradually get better estimates of the state sequence. This framework leads to popular methods like the Kalman filter (for simple, well-behaved cases), the particle filter (for more complex problems), and the histogram filter, among others. Which method you choose depends on the problem and often requires experience.

Sometimes the goal is to estimate the state sequence; other times you might try to learn the behavior of the noise itself. A useful idea is the innovation sequence—the new information each measurement adds—which helps design stopping rules and understand how measurements relate to each other.

Setting the initial conditions for the models can be tricky and is often based on prior data or measurements done under a different setup. Practical filters have many parameters and details, and they may not be reliable for safety-critical applications unless they’re well tested and justified.

If the problem involves a larger context, like mapping an environment while tracking position, techniques called SLAM (simultaneous localization and mapping) are used. Sequential estimation is a core part of SLAM, with the whole system treated as a special case when the overall state is just a single variable.

There are non-causal variants that use all measurements at once or reverse time for batches or post-processing. These are not real-time capable unless large buffers are used, which slows things down. Some methods also run several passes to get an initial rough estimate and then refine it, a bit like editing video. In image processing, where all pixels may be available at once, these methods can be made causal again.

Sequential estimation underpins many well-known tasks, such as Viterbi decoding, convolutional codes, video compression, and target tracking. Its state-space view, often aligned with physical laws, also connects to control theory and has led to the use of filters like the Kalman filter in space-related applications.