Nexus Events is a turn-based strategy game where foresight and chain reactions determine the winner. Players begin their journey from a Hall of Fame screen, which honors past champions. From the main menu, you can access tutorials to master the rules, customize the game's visual theme, and toggle sound effects to your liking. Starting a new game is simple: select your preferred grid, with options to even upload a custom image for a unique battlefield. Next, choose the number of human and AI players, enter your names, and get ready for a strategic showdown. The core mechanic is intuitive yet deep. On your turn, you select a single cell on the grid. This action fills the cell with your color. When a cell reaches its capacity—determined by the number of adjacent cells—it unleashes a chain reaction, bursting open and converting all neighboring cells to your color. This can trigger a cascade of bursts across the board, leading to dramatic shifts in territory control. The game continues with this dynamic, recursive process until one player successfully captures every cell on the grid. When the game concludes, a quick press of the "R" key resets the board for a fresh challenge.

To enhance the single-player experience, I focused on creating a challenging and intelligent AI opponent. The development process was iterative, beginning with an exploration of different machine learning approaches.

My first attempt involved a Convolutional Neural Network (CNN). The idea was to have the model "see" the game board as an image, with different colors representing player territories, and predict the best cell index to click. However, this approach struggled. The CNN had difficulty grasping the abstract, strategic nature of the chain-reaction mechanic and wasn't able to make consistently accurate predictions. A significant hurdle was the inability to generate a valid, comprehensive dataset that could effectively teach the model the nuances of the game's strategy.

Recognizing the limitations of a data-driven approach for a game with evolving strategies, I pivoted to Reinforcement Learning (RL). I wanted an NPC that could learn by playing the game, not just by analyzing static examples. This led me to build a system where two RL agents competed against each other thousands of times. Because RL excels at developing long-term strategies, the agents began to uncover unique and highly effective techniques—patterns and moves that a human player might only discover after months of trial and error. This self-play approach created a truly formidable AI that could challenge even experienced players.

The AI's brain is a Q-network, a neural network built with TensorFlow that approximates the optimal action for any given game state. The game state itself is converted into a numerical vector, capturing the value (number of fills) and color of every cell on the board. This vector is fed into the Q-network, which consists of two hidden dense layers with 128 neurons each, using the ReLU activation function. The final output layer is a dense layer with a linear activation that produces a Q-value for every possible action (i.e., every cell on the grid).

During training, the agents use an epsilon-greedy policy to balance exploring new moves with exploiting known winning strategies. For each move, the agent receives a positive reward for continuing the game and a negative one for a loss, incentivizing it to prolong play and dominate the board. These experiences—state, action, reward, and next state—are stored in a memory buffer. The agents learn by sampling batches from this memory to update their Q-networks, minimizing the difference between their predicted Q-values and the actual outcomes. Over many episodes, the weights of the models are saved, allowing them to build upon their learned strategies and become progressively more intelligent and competitive opponents.