Our menu features two distinct options:

• Normal mode: the user moves the character using arrow keys.
• AI mode: the character is moved using pre-trained reinforcement learning.

How to play in normal mode?

• Once the player clicks normal mode, the game starts immediately. The character begins at the starting square.
• Move the character using arrow keys while avoiding holes.
• Use teleporters to reach the finish square.

How to play in AI mode?

• Once the player clicks AI mode, the game starts immediately.
• Watch the character navigate toward the finish square.

Here are the structures used:

Teleportation Movement

As explained earlier, we must guide our character from the starting square to the finish square via teleporters. The teleporters are linked in pairs by color. When the character lands on one, an animation transitions them to the corresponding teleporter.

Character Death

If our character hits a hole, they are reset to the starting position. We use the teleportation animation; essentially, death teleports the player back to the start.

Character Victory

When the character reaches the final square, they perform a small jump and spin.

Our perceptions are based on the character's X and Y positions on the map. We have 529 states depending on 4 directions.

The AI follows the Q-Learning algorithm. Thus, our QTable is a 529x4 matrix.

Our reward system is simple: 1 if the character reaches the finish point, 0 otherwise. To determine the $\xi$ and $\gamma$ parameters, we performed various simulations over 500,000 generations. The best results were $\xi = 0.8$ and $\gamma = 0.93$.

Action selection follows the E-greedy algorithm. Our epsilon is a geometric sequence with a 0.999 parameter. For our final model, this epsilon decreases every 150 generations.

Once this modeling is complete, we save the resulting QTable and load it into the main program to move our character using the experience from previous generations.

Our finish point corresponds to state 265; states 264 and 266 correspond to the squares to the left and right of the final square.
You can clearly see the 1s in the RIGHT and LEFT columns for each state.

Here is a demonstration of the game in AI mode:

Future Improvements:

• Improve input handling for normal mode.
• Add an object to find to unlock the exit.