Description

This project contains my implemenation of reinforcement learning algorithms to attempt to play the game of clash royale. Implemented ideas from this paper https://arxiv.org/pdf/1803.10122.pdf and https://arxiv.org/abs/1810.06394

Agent Overview

action space: card choice: Discrete(5)

• 0 - NOOP
• 1 - 4 (corresponding card in hand)

position: game screen boundary

Observation space:

• Latent features: shape of z_dimension of vae
• card features: current shape = 166, changes as development continues

Methodology

Card features include:

• current time
• current elixir
• current princess towers hitpoints and damage per second
• 2x elixir (bool)
• XY positions of enemy troops on the board (max of 50)

Then for each card in hand and next;

• card type (categorical, Troop | Spell)
• damage per second
• hitpoints
• hitspeed
• targets (categorical, Air | Ground | Air & Ground | Building)
• speed (categorical, Slow | Medium | Fast | Very fast)
• range (mixed but treated as categorical)
• count
• crown tower damage
• area damage
• radius
• elixir cost

References

• https://github.com/davidADSP/GDL_code
• https://github.com/lollcat/RL-Process-Design
• https://arxiv.org/abs/1511.06581
• https://arxiv.org/pdf/1803.10122.pdf
• https://arxiv.org/abs/1810.06394
• https://www.ijcai.org/proceedings/2019/0631.pdf