kicking the tires on jev (TypeSafe's System One model) with 2048

I recently had the opportunity to test jev, specifically version 1.13.0, with the objective of evaluating its performance in the game 2048. My approach involved providing jev with the current board state and allowing it to choose between four possible moves: up, down, left, or right. To track the results, I ran this test four times, each time corresponding to a different row in the accompanying table.

In the initial trials, jev's performance closely resembled that of a random move generator. However, when the code was able to predict the board's state after each possible move, jev's performance significantly improved. It matched the efficiency of a fixed rule-based system, which could decide on the optimal move depending on the current board configuration.

For instance, if moving left would shift any tiles, it would choose left; otherwise, it would opt for down, then right, and finally up if those options were available.

Interestingly, there were instances where jev's initial move did not alter the board at all. For example, attempting to move left was futile when every tile was already against the left edge of the board, leaving no room for any possible merges. In such cases, jev would then proceed with its second-choice move.

Comparatively, when I employed random moves or adhered to predefined fixed rules, jev's performance was inferior. The random approach, unsurprisingly, resulted in erratic outcomes, while the fixed rules performed consistently based on predetermined strategies. In the final row of tests, the fixed rules were the sole viable option, as the other three techniques had no feasible moves to suggest.