Mind Game · Strategy, AI and behavioural modelling
Mirror
You are not playing against the AI. You are playing against what the AI believes you will do.
A colony of 100 learning agents shares four regions. You never control any of them: you shape their world with one intervention per season. While you try to read them, each agent is quietly building its own model of you, and acting on it.
A new colony
100 agents have settled across four regions. Some cooperate and share; others compete and grab twice as much, which wastes the land. Pick an intervention and a region.
The colony
Your moves so far
You believed you were teaching the agents how to behave.
In reality, your behaviour became their training data.
Their forecast of your next move, season by season
Each column is the colony's average prediction made before you moved. Darker means they were more sure (the scale runs from the lowest to the highest forecast in your game). Hover a cell for the numbers.
How well they read you
Probability the colony gave your actual move (intervention and region). The baseline is chance: 1 in 16 (6.25%).
How the colony changed
Share of agents cooperating, and how random their choices were (100% = completely unreadable).
How they think you choose a region
Their profile of you
Show all seasons as a table
How it works
A social dilemma
In each region, competitors grab twice a cooperator's share, but competition wastes the land. Grabbing pays for one agent and hurts everyone together, as in many real commons.
They model you
Each agent learns, only from your past moves, what you tend to do next (a Markov model with forgetting) and which rule you use to pick a region: the crowd, the competitors, the richest, the poorest, the same as last time...
They act on it
Before your move lands, every agent picks a region and a stance to maximise its expected payoff under its forecast of you. If they expect an audit in Riverside, its competitors quietly move elsewhere.
They hide
Agents caught out while doing the obvious thing, or who watch a big catch, make their own choices more random. Exploit their predictability and they learn to become unpredictable.
The idea behind it
- This is a toy version of theory of mind and opponent modelling: two learners modelling each other. Being predictable makes your moves cheap to counter, which is why randomising (a mixed strategy, in game-theory terms) is often optimal.
- Each agent is different: how fast it forgets, how much it trusts its model of you, and how strongly it believes cooperation pays. So the colony's forecast is the average of 100 slightly different minds.
- Every game uses a new random seed: different land, different agents, and a random event part-way through.
Everything runs in your browser in plain JavaScript; no data is sent anywhere.