Cog_tasks_RL_agents

Create cognitive tasks for Reinforcement Learning agents and benchmark them

Background

Cognitive neuroscientists run a number of experiments in the lab to probe animal and human behaviour. But, machine learning / reinforcement learning (RL) researchers use very different benchmarks to evaluate their learning agents.To make it easier to compare the behavior of animals / humans with these agents, we need to implement the cognitive neuroscience tasks in environments that are accessible to artificial reinforcement learning agents.

What is known:

• The performance of machine learning agent on machine learning task
• The performance of cognitive agent on cognitive task

What is unknown:

• The performance of machine learning agent on cognitive task
• The performance of the cognitive agent on machine learningtask.

Usage

All agents inherit from the basic Agent class in agent.py. If you want to use the agents to train on any of the gym environment, please see the example.py.

Agents

6 agents are implemented in this project:

• AuGMEnT
• LSTM
• DQN
• DRQN (DQN + LSTM)
• HER
• Monte Carlo

Tasks

Implemented in the OpenAI gym style. They are put in a independent repo here.

• 12_AX
• 12_AX_S
• AX_CPT
• 12_AX_CPT
• Copy
• Copy_repeat
• Saccades
• Sequential Prediction

Benchmark

Every agent is trained and evaluated on each of the tasks.