Does evidence reversal change the observed data?

No. It changes how the model is organized for inference.

Does a reversed computational arrow prove reverse causation?

No. Probabilistic refactorization is distinct from a causal intervention claim.

Is it the same as resampling?

No. Evidence-informed proposal generation and weight-based resampling are separate steps, although a filtering method can use both.

AI Glossary / Graphs and symbolic AI / Intermediate

Evidence reversal

Definition

Evidence reversal is a probabilistic inference technique that restructures a model so observed evidence can guide the generation of hidden-state samples. In dynamic Bayesian-network filtering, a common form samples the next state conditional on both the previous state and the new observation. Here, “reversal” concerns the computational direction of probabilistic dependencies, not changing or contradicting the evidence.

What it means

A simple forward sampler first predicts a hidden state from the transition model, then evaluates how well that state explains the observation. If the observation is informative, many predicted samples may receive very little weight.
Evidence reversal incorporates the observation into the distribution used to propose the new state. The technique uses probability identities to represent the relevant distribution differently. It does not claim that an observation physically causes an earlier state, and it does not justify reversing arbitrary arrows while leaving probability tables unchanged.

How it works

For a fixed previous state x_old, combine the transition probability P(x_new given x_old) with the observation likelihood P(e given x_new). Normalize their product to obtain P(x_new given x_old, e), assuming the observation has positive predictive probability and the stated model factorization applies.
That conditional distribution can generate new-state samples already informed by e. The predictive evidence probability P(e given x_old) remains relevant to weighting previous-state particles in a particle-filter formulation. Computing these quantities may be expensive in a large factored model, and reversal can introduce dependencies. The method is therefore a computational tradeoff, not free access to exact filtering.

Three practical examples

A teaching particle filter proposes a hidden machine state using the latest simulated sensor reading as well as its transition model.
A tracking experiment compares prior-based proposals with observation-informed proposals under a controlled probabilistic model.
A graphical-model researcher studies how reversing dependencies toward an observed node changes the intermediate conditional distributions that must be computed.

Think of it this way

Think of selecting plausible explanations after reading a clue instead of generating explanations blindly and discarding most afterward. This captures the purpose of observation-informed sampling. The analogy is limited because a valid probabilistic algorithm must preserve the target distribution through correct conditional probabilities and weights; simply favoring appealing explanations can introduce bias.

Worked example

Fix one previous state in a fictional two-state model. Before seeing the next observation, the transition probabilities are P(A) = 0.8 and P(B) = 0.2. For the observed signal e, let P(e given A) = 0.1 and P(e given B) = 0.9.
The unnormalized conditional weights are 0.8 × 0.1 = 0.08 for A and 0.2 × 0.9 = 0.18 for B. Their sum, 0.26, is the predictive probability of e for this previous state. The observation-informed probabilities are therefore 0.08/0.26 = 4/13 for A and 0.18/0.26 = 9/13 for B.
A prior-based proposal would choose B only 20 percent of the time. The conditional proposal chooses it about 69.23 percent of the time. This is a statement about proposal probabilities, not guaranteed counts in a small random sample. With several previous-state particles, their differing evidence probabilities must also be accounted for; the calculation alone does not justify giving all parent particles equal weight.

Illustrative case study

In a fictional filtering seminar, the Pine team notices that most samples from its transition model fit an unusual observation poorly. It implements a small exact conditional calculation for each previous state and compares the resulting proposal distribution with the prior proposal.
The team then checks weighted estimates against exhaustive enumeration in the tiny model. It reports the extra calculation required per particle instead of treating the more informed proposal as automatically faster. This invented case illustrates verification of an inference method, with no claim of measured improvements in a real tracking system.

Common misconceptions and mistakes

Evidence reversal does not mean retracting a scientific finding when new research disagrees with it. That is another use of the same phrase outside this graphical-model context. Nor is it a causal discovery procedure.
Do not reverse an edge and retain its old conditional table. The new factorization must be computed correctly and may require additional connections. Do not confuse observation-informed proposal sampling with resampling particles according to weights: they address related problems through different operations and can be combined. Finally, a useful proposal does not remove all sampling error or model error.

Business use

For teams evaluating probabilistic monitoring systems, the concept explains one way to reduce the mismatch between generated hypotheses and incoming observations. It is especially relevant when comparing why two particle filters use the same underlying model but produce samples differently.
Ask whether the conditional proposal can be computed within the available budget and whether its weighting has been verified. Compare accuracy and computation under representative observations. The technique changes inference organization; it cannot compensate for incorrect transition dynamics or a sensor model that assigns misleading likelihoods.

Check your understanding

In the worked example, why is the proposed probability of B 9/13 rather than 0.9?

Answer: The likelihood 0.9 must be combined with B's transition probability 0.2 and normalized against A's contribution. The posterior proposal is 0.18 divided by 0.26.

Related terms

Learn next

Learn next about particle-filter importance weights. Repeat the example for a second previous state with different transition probabilities. Compute its predictive evidence probability and explain why two parent particles may deserve different weights even when both use the observation-informed next-state proposal.

Sources

Kanazawa, Koller, and Russell, original UAI 1995 paper, Stochastic Simulation Algorithms for Dynamic Probabilistic Networks: https://arxiv.org/pdf/1302.4965 Original paper bibliographic record and abstract: https://arxiv.org/abs/1302.4965

Previous: Evidence Next: Evidence upper bound