Extended Data Fig. 4: Double T-Maze evolved learning behaviour.

Double T-Maze evolved learning behaviour. Example of steps taking in the double T-maze environment by an evolved Network of ENUs. The agent can be seen to have successfully evolved to explore the environment to find and eat the initial poison (1). It then explores an alternative path to find non-poisonous food instead (2), indicating it has properly learned from a single example to associate the previous actions taken with a negative reward. Since food and poison can randomly change location, the agent goes back to the previous food location, but detects poison instead. As it previous obtained a negative reward with the action of eating the poison, it internally modified the synapse ENUs internal memory states to alter its behaviour, and successfully learned to turn around and find food in another part of the maze (3). It also evolved proper exploration behaviour if no food or poison is found in a section of the maze, successfully navigating to the other side (4).