17. Evolving a Dota 2 Hero Bot with a Probabilistic Shared Memory Model

Reinforcement learning (RL) tasks have often assumed a Markov decision process, which is to say, state information is ‘complete’, hence there is no need to learn what to learn from. However, recent advances—such as visual reinforcement learning—have enabled the tasks typically addressed using RL to expand to include significant amounts of partial observability. This implies that the representation needs to support multiple forms of memory, thus credit assignment needs to: find efficient ways to encode high dimensional data, as well has, determining under what conditions to save and recall specific pieces of information, and for what purpose. In this work, we assume the tangled program graph (TPG) formulation for genetic programming, where this has already demonstrated competitiveness with deep learning solutions to multiple RL tasks (under complete information). In this work, TPG is augmented with indexed memory using a probabilistic formulation of a write operation (defines long and short term memory) and an indexed read. Moreover, the register information specific to the programs co-operating within a program is used to provide the low dimensional encoding of state. We demonstrate that TPG can then successfully evolve a behaviour for a hero bot in the Dota 2 game engine when playing in a single lane 1-on-1 configuration with the game engine hero bot as the opponent. Specific recommendations are made regarding the design of an appropriate fitness function. We show that TPG without indexed memory completely fails to learn any useful behaviour. Only with indexed memory are useful hero behaviours discovered.