We will generate a start and end coordinate in the map. The starting point will be on ground level. The ending point being some coordinate 100-1000 blocks, not including verticality, away from the starting point. Our agent, Steve, wil spawn at the start coordinate and has a goal of reaching the end coordinate. Positive rewards include: arriving at the end coordinate and moving closer to the end coordinate. Negative rewards include: each action taken. Steve will be given a 15x15 array of the relative block heights around himself along with his current and destination coordinates. Movement actions include moving forward, backwards, left, right, and jumping.
We plan on implementing reinforcement learning using a neural network.
We will be evaluating the success of our project by measuring the amount of actions it takes to reach the end coordinate and the distance of the agent from the end coordinate, if it fails to make it. We expect to see improvements from just random actions to movements toward the destination coordinate. The baseline is movement closer to the end coordinate. We will also monitor the agent's pathing by graphings its coordinates on a 2d plane to see how well the agent is doing and if it is improving from just taking random paths. For the moonshot case, we would like to see the agent able to reach the end coordinate through varying terrain.
Wednesdays 11AM