Projects Last Projects

In recent years, the introduction of deep reinforcement learning allowed rapid progress in the pursuit after implementing general AI. One of the long-standing challenges withholding further progress is designing an agent that operates in a hierarchical manner with temporal abstractions over its actions. We present a system which disassembles the learning into multiple sub-skills without external assistance. The system consists of a deep recurrent network which learns to generate action sequences from raw pixels alone, and implicitly learns structure over those sequences. We test the model on a complex 3D first person shooter game environment to demonstrate its effectiveness.

Please, see project report.

Please, see final presentation.

Please, see data set.

Please, see Code.