Introduction: Our group tried to find an optimization strategy to play the game called Jstris. This game is an adaptation of tetris with some additional rules such as an ability to attack other players, or some other techniques such as T-spin, following this overview https://jstris.jezevec10.com/guide#overview This problem relates to Reinforcement Learning since we’re interested in making the machine to learn based on the rewards of the score it could get instead of having the specific labels. We may include some use of options such as changing from Deep Q Learning, which is what the research implements into REINFORCE that we study in the class Challenges: What has been the hardest part of the project you’ve encountered so far? : By now, the hardest part so far is including new metrics of calculating our scores. For example, we have to include some scoring system that coordinates with the movement of keyboards such as T-spin. In this case we now consider spotting T-spin using the current state and the next state to see if there is a T-tetromino placed in a more complex way than just a simple drop. Another problem is how many T-spin the agent played since it yields a different score (based on https://n3twork.zendesk.com/hc/en-us/articles/360046263052-Scoring ). In our game, we may simplify it into having one-score-for-all T-spin Insights: Are there any concrete results you can show at this point? : By now, we could come up with the result based on the traditional way of playing tetris. We still cannot make a promising output related to the T-spin action. How is your model performing compared with expectations? : By now it doesn’t work as we expect yet since we could not apply T-spin into the reward model
Plan: Are you on track with your project? What do you need to dedicate more time to? : By now, I think making the new reward function that considers the T-spin and also detecting them are the main priority What are you thinking of changing, if anything? : May change or using a choice between DQN and REINFORCE May simplify the T-spin score for an easier implementation
Log in or sign up for Devpost to join the conversation.