• Language: Scala
  • Rank: #3 (Legend) / 2022
  • Challenge: Troll Farm

Introduction

Early on, a clear meta emerged around chopping trees. A minor rule change later rebalanced things, but I decided to stick with it — chop trees with 2 trolls — since it was straightforward to implement well.

Strategy Overview

My bot follows three broad phases:

  1. Early game: gather just enough resources to train a second troll.
  2. Main loop: both trolls chop trees and return wood, coordinated to avoid conflicting targets.
  3. End game: plant remaining fruits for extra points

Early Game: Resource Objective

The goal was to train the strongest second troll as quickly as possible. I checked whether I could collect enough resources (plum, lemon, iron) to hit the next threshold (5 or 10 of each) in time, based on the round-trip distance to the nearest relevant tree or mine.

Main Loop: Candidate Scoring

The heart of the bot is a candidate generator. For each troll it produces a list of (command, score) pairs and picks the highest-scoring one. The full priority stack:

Score Action
8 000 Drop resources at shack
7 000 Move to shack adjacency (carry full)
dynamic Chop / move to tree

Tree Chopping Value

For every available tree the bot computes a throughput score:

value = min(treeSize, carryRemaining) / (travelTurns + chopTurns + returnTurns)

This is wood per turn: the wood actually collectible divided by the full round-trip cost. travelTurns and chopTurns account for the troll’s movementSpeed and chopPower respectively (using math.ceil for fractional turns).

typeToCut: Picking a Focus Tree Type

On the first turn, the bot picks either LEMON or PLUM as the primary chopping target. The criterion is simple: whichever tree type has its cluster closest to the shack (summed BFS distances across all trees of that type). When a tree matches typeToCut and the opponent has at most 2 trolls, the bot uses a different scoring formula that rewards trees near from the opponent’s shack.

The idea: cut trees that will prevent the opponent from training.

Tree Simulation

A key ingredient for both target selection and safety checks is tree state prediction. The Tree.nextN(n, oppChopPower) method simulates n turns of growth and opponent chopping. The bot uses this to:

  • Skip trees that will die before the troll arrives (opponent is already chopping it).
  • Compute the actual health/size at arrival time to get accurate chopTurns and wood yield.

The opponent’s chop power is inferred from the isDamaged flag on the tree: if the tree’s health differs from its expected value for its size, someone is chopping it.

Multi-Troll Coordination

With two trolls the bot enumerates all pairs of candidates (a, b) where a and b do not target the same cell, then picks the pair with the highest combined score.

“Same target” covers moves toward a tree, chopping, harvesting, and dropping at the shack.

End Game

When trees are almost gone and the opponent leads, or when turn > 250, the bot switches to end-game mode:

  • Trolls pick fruits from the shack and plant them on empty cells for extra scoring points.
  • Trolls carrying resources rush to the shack to drop.
  • The fallback waiting position shifts to a cell adjacent to the opponent’s shack, staying in position to contest any last-minute planting.

Things to Improve

Sweet spot (disabled)
The sweet spot feature (planting an apple tree on a cell adjacent to the shack and next to water for a very fast fruit cycle) was implemented but disabled

Movement optimisation I didn’t optimize movement at all. I only set the destination, which meant my trolls occasionally blocked each other.

The competition inside the competition

My son m4l0s4n also competed, doing remarkably well in Python 3. He reached Legend before me and finished #66. I couldn’t be prouder. We had a great time together, discussing strategies and watching replays. Every time he overtook me in the rankings, I had no choice but to push harder and come up with new ideas. He made sure I knew it too, with his profile tagline:

son of yamo but better

Conclusion

Thanks a lot to eulerscheZahl for such a fun challenge. I really enjoy heuristic-friendly challenges. To be honest, I’m really surprised to see how far this simple idea got me in the rankings.