[All work](/work?type=work)

# Sentient Environment Explorer

Sentient was an experiment in an agent learning its way through a changing 2D environment, with food and hazards giving its movement a consequence.

Independent experiment development. Python, deep Q-learning and simulation UI.

![Explanatory diagram for Sentient Environment Explorer, showing Environment, Agent action, Reward and replay.](https://cdn.sanity.io/images/qnuj1c4o/production/b82155af4c5950a534655d4928065f3a380fa822-1280x720.png?w=1280\&q=80\&auto=format)

I built a small world where movement mattered to the agent's survival, with food, harmful objects and a hunger mechanism giving it more to do than simply cross an empty grid, with the environment changing as the agent moved and the reward values shaping which actions the learner would revisit.

The code separates the environment, neural learner and orchestrator, with a GUI for pausing the simulation, changing its speed and observing the behaviour while learning continues.

The repository includes saved checkpoints from training runs, though those show that the learning workflow ran and do not by themselves establish a general improvement in exploration or decision making.

The catalogue date follows the first preserved commit on 10 March 2025.

## Outcome

A visual reinforcement-learning prototype with a running environment, controls and saved model checkpoints. The public cover is a schematic world, not a frame from a recorded run.

[All work](/work?type=work)