Collaborative modelling by BayesFusion
Experts, modellers and data scientists work on one live model in the browser. Every change is recorded with who made it.
What changes
BayesTeam turns a Bayesian network from a file one modeller owns into one shared, live model the whole team builds and questions together, with every change recorded.
BeforeThe model lives in one person's file, and versions travel by e-mail.
With BayesTeamOne model in the browser. Several people edit it at once, and each sees who is in it and every change as it is made.
BeforeNobody knows who changed which probability, or when.
With BayesTeamEvery change is recorded with its author and time, and the Activity Log lists them, newest first.
BeforeStakeholders cannot look at the model without installing and learning a tool, so every question goes through the modeller.
With BayesTeamInvited colleagues open the model in a browser. With the Viewer role they follow every change, set evidence and read each node's values, without being able to change the model.
BeforeReviewing a case together means watching one person's shared screen.
With BayesTeamEvidence is shared by everyone on the model, so the team discusses the same beliefs, with the team chat and an audio call beside the model, and keeps scenarios as cases.
BeforeData and expert knowledge live in separate tools.
With BayesTeamData sets, fitting the tables to data, structure learning and validation work on the same model the experts edit.
BeforeGetting started takes a long time.
With BayesTeamAn AI assistant drafts edits from plain language for a person to apply, and narrated video guides are built into the app.
The idea
A map of what influences what, with numbers attached. Tell it what you have observed and it works out, exactly, what that means for everything you have not. The network below is drawn the way BayesTeam draws it.
Each node is something uncertain: is it cloudy, is the sprinkler on, did it rain, is the grass wet? Each has a few states, and in the bar chart view every state gets its probability and a bar.
An arc from Cloudy to Rain says clouds change the chance of rain. There is no arc from Cloudy to Wet grass: clouds matter to the lawn only through the sprinkler and the rain.
Its definition gives the chance of each state for every combination of the parents' states. The grass is wet with 0.99 when the sprinkler and the rain both do their work, 0.9 with one of them, and never with neither.
Lawn
You
Node Definition: Wet grass
| Sprinkler | on | off | ||
|---|---|---|---|---|
| Rain | yes | no | yes | no |
| yes | 0.99 | 0.9 | 0.9 | 0 |
| no | 0.01 | 0.1 | 0.1 | 1 |
Click a state on a node or in the Nodes list to observe it, or pick a question below. The numbers are computed exactly from the tables, and printed as BayesTeam prints them.
From cause to effect. On a cloudy day rain becomes likely (80.00%), the sprinkler rarely runs (10.00%) and wet grass rises to 74.52%.
From effect to cause. The grass is wet, so something wetted it: rain goes from 50.00% to 70.79%, the sprinkler from 30.00% to 42.98%.
Explaining away. Now we also learn that the sprinkler was on. It already explains the wet grass, so rain falls back from 70.79% to 32.04%, even though the sprinkler and the rain do not influence each other directly.
The platform
The model lives on the server and everyone works on the same one, in the browser. The pictures below are drawn from the app's own panels, step by step as they behave.
Working together
Everyone who has the model open works on the same network and sees what the others do, as they do it. Here Alex watches Sam work; scroll to follow it step by step.
Lawn
SD
Alex Demo
Activity Log
Team chat · Lawn
Sam Demo
I raised Wet grass when it rains and the sprinkler is on. Can you check?
Type message
Step by step
Evidence and cases
Evidence is shared by everyone on the model. With Update Immediately on, every change is answered at once; here it is off, so each stage can be seen.
Lawn
Alex Demo
Case Manager · Lawn
ApplyFilter
| Name | Category | Description | Wet grass |
|---|---|---|---|
| Wet morning | Garden | The lawn is wet at dawn | yes |
Step by step
Diagnosis
Give nodes a diagnostic role, faults or observations, and run the diagnosis. The numbers are those of the Flu clinic model in the video guide.
FaultsProbability
Ranked ObservationsDiagnostic Value
Other Observations
Past ObservationsState
Step by step
AI assistant
Ask for a change in plain words. Nothing reaches the model until someone applies it. Your administrator decides whether the assistant is switched on.
Lawn
Alex Demo
Assistant
Add a node called Season with states dry and wet, and make it a parent of Rain.
Added the Season node and connected it as a parent of Rain.
AppliedAdd node (with connections & table) Season
▸ 2 atoms
states [dry, wet]
AppliedAdd connection Season → Rain
+ Season → Rain
▸ 1 atom
5 nodes, 5 arcs after the change.
Apply 3 changesDiscardWhy? (optional)
Applied to the model.
Ask the assistant...Ask
Step by step
Who it is for
Typical uses are diagnosis (medical, technical or fault diagnosis, with faults and observations ranked by diagnostic value), risk and reliability models, and decisions under uncertainty with influence diagrams.
The real thing
Recordings of BayesTeam itself, from the narrated video guides that come with the app. Short excerpts, muted.
Sam opens the same model: a second face in the navbar. Sam selects Rain: a ring in Sam's colour. Sam double-clicks Wet grass: the editing badge.
View Nodes As Bar Chart, turn on Update Immediately, then double-click a bar: it becomes the evidence and every bar moves.
Double-click a case: its evidence is applied. Change a cell, update, and the applied case takes the new evidence in.
Run the diagnosis: faults and questions are ranked. Pursue a fault, answer a question, and the ranking moves.
Ask for a new node in words. The assistant reads the model, plans, and lists each change; Apply puts them into the model.
Features
Everyone on a model sees who else is there, what they have selected or are editing, and every change the moment it is made.
Every change is recorded with who made it and when, newest first, and a group of changes opens to show what it held.
A chat beside every model, kept with it, and an audio call when typing is too slow.
Observe a state and update: every node shows its new probabilities. Probability of evidence and the most probable explanation are on the toolbar.
The Case Manager keeps sets of evidence inside the model; double-click a case to apply it.
Mark faults and observations, then rank the questions by diagnostic value. Pursue a fault, answer a question, rank again.
Decision, value and MAU nodes beside chance nodes, with value of information for a decision.
A temporal plate with a number of time steps, temporal arcs, evidence per step, and unrolling.
Equation nodes for continuous variables, and Noisy-MAX, deterministic and DeMorgan nodes.
Upload a data set, fit the tables to it, learn a structure, and validate the model.
Ask for a change in words, read the changes it proposes, then apply or discard them. Switched on by your administrator.
Short narrated guides in the Help menu, recorded in the app itself.
Questions
A Bayesian network is a model of uncertain variables and the direct influences between them. Each variable is a node with a few states, each arc is a direct influence, and each node has a table of probabilities for its states given its parents. When you observe some variables, the network computes how likely the states of all the others are.
No. BayesTeam is a product of its own, developed alongside BayesFusion's other tools, for teams that build and question a model together in the browser. It reads and writes XDSL, so existing models open in it as they are and models leave it as XDSL.
Yes. Every change is sent to everyone who has the model open and recorded with its author and time. A colleague's selection and editing are marked on your canvas in their colour. If two people save the same table at the same moment, the first save stands and the second is refused, so nothing is overwritten unseen.
Models are stored on the BayesTeam server and computed there. Only accounts an administrator has added can sign in, with the role of viewer, editor or administrator. An organisation can run its own BayesTeam server on its own machine. A cloud compute service from BayesFusion is optional and used only when the administrator configures it.
BayesTeam reads and writes XDSL, BayesFusion's model format. Upload an existing .xdsl file to work on it with your team, and download any model as XDSL.
Access is by invitation. To request access for yourself or your team, write to support@bayesfusion.com. Once you are added, sign in with your Google account.
Yes, where the administrator has switched it on. You ask for a change in plain words; the assistant reads the model, plans the edit, replays it on a copy to check it and lists every change. Nothing changes until someone applies the proposal. You can use your own key for the language model, kept in your browser.
Exact algorithms (clustering, relevance decomposition, polytree) and sampling algorithms (logic sampling, self-importance, heuristic importance, backward, AIS and EPIS sampling), chosen per model. BayesTeam also computes the probability of evidence, the most probable explanation, value of information and diagnostic rankings.
Yes. Upload a data set, prepare it, then fit the tables of an existing model to the data or learn a new structure from it, and validate the model against the data.
Access
To ask for access for yourself or your team, write to support@bayesfusion.com. Already invited? Sign in with the Google account your administrator added.