A Procedurally Generated Platform

First things first, we need an environment for our agents to learn in.

Dynamic Platform, Dynamic Challenges

At the heart of our experiment lies a platform composed of hexagonal blocks that dynamically adjust in size. This feature isn't just for visual flair; it serves a crucial role in challenging our AI agents to adapt to ever-changing landscapes. As agents progress through the game, achieving three consecutive successes triggers an increase in both difficulty and platform size. This scalability aims to test the agent's adaptability and decision-making capabilities under varying conditions.

The platform can scale its area, to create an increasing level of difficulty.

The Hunt for the Moving Goal

Central to our environment is a randomly placed goal that serves as the primary objective for our AI agents. Upon touching the goal, agents are rewarded, reinforcing successful behaviors. This setup encourages agents to navigate efficiently and strategically across the dynamically shifting terrain without overfitting to a goal that is always spawned in the same location.

The goal's position is randomized, to help the agents generalize it's learned behaviours.

Timing and Consequences

To add a layer of urgency and consequence, we've implemented a countdown timer that dictates the duration of each game session. If the timer expires or if an agent falls off the platform, penalties ensue, challenging agents to balance speed and accuracy in their decision-making.

Progressive Learning Curriculum

Our platform is not just about immediate challenges; it's designed as a progressive curriculum. Starting with a small, manageable platform and goals placed within easier reach, the environment evolves as agents demonstrate proficiency. This evolution mirrors real-world learning scenarios, where increasing challenges correspond to growing competence.

But Wait, There’s More!

We went crazy… We aren’t limited to the platform just expanding around its perimeter; it can generate a variety of shapes! We've developed multiple algorithms to create even more challenging environments for our agents. Do you think they can handle it?