A talk about Automating discovery

James Hetherington

2026-07-22

Let’s learn a game

Why isn’t chess like science?

• You know where all the pieces are.
• You know what the valid moves are.
• You know what counts as a win or a draw.

How to play Kriegspiel

For Kriegspiel, you need three boards in three rooms

The referee has the only board with the real positions of all the pieces.

How to play Kriegspiel

Suppose the position is this:

…I play Qd8.

The referee tells me “illegal move, there is a pawn on d5”.

What does Kriegspiel tell us about science?

• We can play to try to achieve some engineering objective
• But we also play to learn about the world

• We evaluate a move by how much it reduces uncertainty, as well as whether it achieves the objective.
• We call this a “partially observed decision problem” (PODP).

Simulating gravity

Simulations and models help us write down our current view of how the world works.

We can make simulators of gravity:

Physics-free gravity

• We can observe lots of pictures of planets.
• Try to learn the rules of gravity from those observations.

The triumph of stamp collecting.

Image credit: Wikimedia commons, Heptagon, CC-SA.

“All science is either physics or stamp collecting” - attributed to Ernest Rutherford, perhaps apocryphal.

Uncertainty in physics-free gravity

What if we could move the planets?

Automated experimentation