Statement
The Interacting Multiple Model (IMM) algorithm of Blom & Bar-Shalom (1988)
achieves tracking accuracy statistically indistinguishable from the
second-order Generalized Pseudo-Bayesian (GPB2) estimator on standard
maneuvering-target benchmarks while running only r Kalman filters per step
— matching the per-step cost of the much weaker first-order GPB (GPB1)
estimator. The mechanism is the IMM interaction step, which forms r
Markov-weighted mixed initial conditions at the start of every cycle, giving
each mode-conditional filter an effective two-step memory of the mode history
without explicitly carrying r² filters as GPB2 does.
Evidence summary
- mazor-imm-target-tracking-survey (1998 IEEE TAES, importance 4): survey consolidates Monte Carlo studies from the decade following Blom & Bar-Shalom (1988) on canonical 2-mode and 3-mode maneuvering-target benchmarks. IMM RMSE matches GPB2 within MC noise; GPB1 at the same cost loses noticeably during mode transitions. By 1998 IMM is the de facto production standard for maneuvering-target tracking in operational radar systems.
Conditions and scope
Conditions are listed in the frontmatter conditions field. Briefly: linear
Gaussian mode-conditional dynamics, small known mode set, known time-invariant
transition matrix, and likelihoods that depend on at most a two-step mode
history. Outside these conditions (nonlinear/non-Gaussian filters,
long-history-dependent likelihoods, unknown mode set, large r), IMM still
runs but the moment-matching bias is no longer benign and methods like RBPF
or full particle filters become competitive.
Counter-evidence
None recorded yet in the wiki. The Mazor 1998 survey itself notes that the moment-matching approximation degrades for highly skewed or multi-modal mode-conditional posteriors, but does not provide a benchmark on which IMM is shown to fall meaningfully short of GPB2.
Linked ideas
(None yet.)
Open questions
- Does the IMM ≈ GPB2 equivalence hold when mode-conditional dynamics are nonlinear (EKF-IMM, UKF-IMM)?
- Does the equivalence survive when the Markov transition matrix
Πis estimated online rather than known? - For problems where the likelihood depends on the full mode history (e.g. asset-pricing models with Riccati recursions over the regime path), how much accuracy does IMM lose compared to a Rao-Blackwellized particle filter that carries the full mode history per particle?