Hi, I finished around 1200th. Hardly a good or “achievement/T-Shirt” rank, but I want to tell yo a few words.
Personal record:
For first time I’ve reached the Gold League in a contest (I’ve participated in around 7 contests before that) and I’m proud of myself 
Lessons learned:
As always I’ve learned several things throughout the contest, which I’ll definitely try to follow the next time:
- Implementing a heuristic(simple if else logic) bot, based on the current game state, without looking in the future is very very important. Until now I wasn’t a big fan of heuristic evaluations, pruning moves and learning/analyzing the game myself, I just wanted to implement a general bot, given the state to output a move.
- But this time I’ve decided to start with a heuristic bot and it was worth it. It was fun, just chilling morning/afternoon programming, no fancy algorithms, no complex tree/graph data structures, no messing probabilities a dozen times, … only observing and trying new things. This really got me into the game and I was more and more eager to move forward.
- For first time I’ve watched a streamer and saw him analyzing a game of the top players very deeply (something which I don’t do), which helped me coming up with new ideas myself. Also chatting with outer players was nice.
- This helped me made a bot which SEEDs only when the seed cost is 0(on the best soil), COMPLETEs lately in the game and GROWs the trees on the best soil. And that’s it, this several lines of code moved me to SILVER league. In my opinion a heuristic bot could get you in GOLD, so dare to try
- After that I was ready to make a simulation of the game with given commands sequence and look into future.
- The simulation part is very very important for any kind of AI approach, so implementing fast and correct simulation is a must.
- I was wondering how to test my simulation and came up with the idea to try classic Monte Carlo simulation approach, get the possible moves for the current turn, play as much random games from each one until the end as the time allows, choose the turn with most wins.
After fixing simulation bugs for 2 days I saw that my MC couldn’t reach a lot of simulations, but after incorporating the ideas from my heuristic bot the simulations increased and I was able to reach GOOOLD
- Then I wanted to implement MCTS and almost got it, but unfortunately couldn’t polish it for submit, which could got me in LEGEND, I’ll definitely try it in the classic game later.
- Time management is also very important, for first time I spent at least 3 hours per day for each day of the contest, which definitely paid off.
TODO:
- I want to optimize my simulation and available commands gathering, maybe by using bitboards
- Implement MCTS with random playouts
- Implement MCTS with heuristic bot playouts
- Implement MCTS with evaluation “playouts”
- Implement MCTS with solver
The challenge was very fun