Train the agent to win exciting prizes at BASALT Minecraft Competition

Deep reinforcement learning combines reinforcement learning and deep learning. It implies a simple rule for reward and penalisation. Deep reinforcement learning assists in the training of software agents. This subfield of machine learning has become testing sites for a few games namely Atari, Chess, and Rubic’s cube. 

 

To accelerate research in this expansive field, Researchers at DeepMind and the University of California have organised a competition named BASALT Minecraft. There will be no pre-specified reward functions in BASALT. The prime aim of the competition is to train an AI system that it communicates via demonstrations, preferences and human feedback. 

 

BASALT is a combination of a minecraft environment and human evaluation protocol. Systems in BASALT are required to learn and understand intricate details of specific tasks through human feedback. BASALT will replicate realistic settings. 

 

(Must Check: What is Game Theory in AI ? Nash Equilibrium )

 

BASALT will specify tasks for designers. It will also enable them to develop the agent and solve the task with no holds barred. Participants will be given a Gym environment along with task description. Gym environment will showcase relevant details like pixel observation and player’s inventory. Designers will use feedback modalities and hardcoded heuristics for creating agents. 

 

Evaluations will be done based on human comparisons, trajectories of two different agents will be recorded in the same environment to simplify comparisons. Later human game players will choose the best performing agent. The code that will be used in training gameplay agents will be later released. 

 

After multiple evaluations the model will be retrained on winning code, to compare the working system of submitted agents. 

 

Important Dates

 

• October 15, 2021: Last Date for Entry Submission

• November, 2021: Winners will be announced through Minecraft’s official website

• December 13, 2021: Winners will be given $ 5000.  Code will be released, presentation of code at NeurIPS 2021

According to VentureBeat Rohin Shah Project Lead at BASALT said, “We envision eventually building agents that can be instructed to perform arbitrary Minecraft tasks in natural language on public multiplayer servers, or inferring what large-scale project human players are working on and assisting with those projects while adhering to the norms and customs followed on that server.”