Is the prior learned from the task data?

The fixed prior function is not; the trainable correction responds to those data.

Does each prediction require new random weights?

No. Each member retains its sampled prior, supporting consistent predictions across inputs and time.

Is the ensemble an exact Bayesian posterior?

Not generally. Exactness requires particular model and randomization assumptions, rather than the presence of a frozen network alone.

AI Glossary / Probabilistic machine learning / Advanced

Randomized prior network

Definition

A randomized prior network is a randomly initialized neural network whose parameters remain fixed and whose output is added to a trainable model. Different ensemble members receive different fixed prior functions, creating initial diversity that can persist where observations provide little constraint.

What it means

Imagine several models that have seen the same small dataset. Agreement outside that dataset may reflect shared assumptions rather than strong evidence. Randomized prior functions introduce differing initial predictions explicitly, while trainable components learn corrections from observations.
The intended uncertainty is primarily about what the model has not learned, rather than the irreducible variability of individual outcomes. In reinforcement learning, a member's value predictions can guide a coherent sequence of exploratory actions. The term here refers to the fixed additive prior construction, not to every network described elsewhere as a prior network.

How it works

For member k, write its prediction as Q_k(x) = f_k(x) + beta p_k(x). The function p_k is sampled once and frozen; f_k is trainable; beta controls the prior contribution's scale. Training evaluates the loss on the combined prediction, so f_k learns to correct the prior where data constrain it. Ensemble members may receive bootstrap samples or other appropriate data perturbations.
For value learning, the prior contribution belongs in the combined value estimate used for both current predictions and member-specific targets. Selecting one member for an episode can preserve consistency across successive decisions. Merely injecting fresh independent noise into every action does not achieve the same behavior.
Randomized additive functions have a precise Bayesian interpretation in particular linear Gaussian constructions with suitable data noise and regularization. Arbitrary deep ensembles do not automatically inherit exact posterior sampling or calibrated uncertainty.

Three practical examples

Simulated routing: an agent considers unfamiliar routes through an environment where useful rewards arrive only after several connected choices. Different members can support different exploratory plans.
Equipment experimentation: an ensemble supplies differing response predictions for operating settings with limited historical coverage. A controlled experiment can investigate disagreement while respecting operating limits.
Regression research: a team compares uncertainty estimates inside and outside the observed input range, separating effects of the fixed prior from effects of random initialization and resampled data.

Think of it this way

Think of several analysts beginning with different provisional sketches, then tracing corrections over them as measurements arrive. Their final drawings include both the starting sketch and the corrections. The sketch is not secretly updated, and the corrections can cancel it where evidence is strong. The analogy does not imply that an arbitrary sketch represents a well-calibrated probability distribution.

Worked example

Consider a deliberately simple two-input model with independently adjustable corrections at an observed location A and an unobserved location B. Set beta = 1. Three fixed priors have outputs at A of -1, 0, and 2, and outputs at B of -2, 1, and 3.
The observed target at A is 4. With unregularized squared loss, fitted corrections at A are 5, 4, and 2. Adding prior and correction gives 4 for every member. If corrections at B remain zero because this toy model has no coupling between locations, predictions at B remain -2, 1, and 3.
Their mean is 2/3. Using the equally weighted empirical distribution, variance is [(-2 - 2/3)^2 + (1 - 2/3)^2 + (3 - 2/3)^2]/3 = 38/9, approximately 4.22. The observed location has zero ensemble spread in this noiseless illustration.
This calculation shows additive cancellation and retained diversity. Real neural networks generalize across inputs, so unobserved predictions need not remain unchanged; three members also do not establish reliable uncertainty calibration.

Illustrative case study

In a fictional project, Juniper Robotics trains a warehouse-navigation agent entirely in simulation. An early ensemble assigns similar low values to every unexplored corridor after repeatedly seeing zero reward near the starting area. The researchers add independently drawn fixed prior functions and train each member's correction on perturbed experience.
They compare complete navigation episodes, failed explorations, and reward discovery across repeated seeds. They also inspect whether a selected member maintains a consistent plan through several junctions. Before any physical trial, the team applies separate collision constraints and tests unfamiliar layouts. This case illustrates an evaluation design, not a reported improvement or permission to explore unsafely.

Common misconceptions and mistakes

The prior network should not be updated by the task loss; otherwise the construction changes. Random initialization of a trainable model is not equivalent to retaining a frozen additive function. Large disagreement is not proof that one member is correct, and agreement is not proof that an unfamiliar input is understood.
Prior scale matters: an overwhelming prior can resist useful correction, while a negligible prior may add little diversity. Ensemble variability is also not automatically observation noise. Finally, averaging all members before every decision can remove the distinct exploratory behavior intended by selecting a member.

Business use

Use this approach when uncertainty affects information gathering and the organization can evaluate exploration carefully. Compare against an ensemble without fixed priors under matched compute, data access, and evaluation conditions. Track task reward alongside costly experiments, constraint violations, and sensitivity to random seeds.
Record the architecture, frozen parameter seeds, prior scale, training perturbations, and member-selection procedure. Check that supposedly frozen parameters truly remain unchanged. Test uncertainty against errors on held-out and shifted inputs rather than treating attractive uncertainty plots as sufficient validation.

Check your understanding

A frozen prior predicts -3 at an observed input whose target is 5. What correction gives a combined prediction of 5 when beta = 1?

Answer: The trainable correction must predict 8, because 8 + (-3) = 5. Training the correction directly toward 5 would instead produce a combined value of 2.

Related terms

Deep ensemble
Bootstrap
Thompson sampling
Exploration
Randomized prior function
epistemic uncertainty
bootstrapped DQN
value function
uncertainty calibration

Learn next

Implement the two-location example, then introduce a shared neural representation and observe how fitting A changes predictions at B. Compare member selection with ensemble averaging in a small simulated decision problem.

Sources

Osband, Aslanides and Cassirer, Randomized Prior Functions for Deep Reinforcement Learning, original paper, Sections 3–4: https://proceedings.neurips.cc/paper/2018/file/5a7b238ba0f6502e5d6be14424b20ded-Paper.pdf Osband, author explanation and additive-model illustrations: https://ianosband.com/2018/11/30/Randomized-Prior-Functions.html Osband-Aslanides-Cassirer fixed additive prior and episode-wise member-selection passages were independently consulted.

Previous: Randomized non-comparative trial Next: Randomized Response