Game Playing Agent

• The agent is trained using proximal policy optimization
• The following is the result after 500,000 training steps

Performance

Trial1.mp4

---

t2.mp4

TODO

• Implement soft-actor-critic (SAC).
• Emperically optimize, neural network parameters: depth, input layer
• Hyperparameter-tuning.

Reward Function

The reward funciton is:

timeSurvived + ( 3 * rocksDestroyed ) + ( 5 * enemyShipsDestroyed )

Rational: the agent would be incentivised to shoot down enemy bullets and rocks and miximize its survival time

Please see this script for more details

Input observations are, positions of objects currently in the scene, feeded sequentially with a label after each position to differentiate between objects

Parameters Used:

• Input layer neurons: 30
• Depth: 2

Shortcomings

• Limited Computational Power