Pax includes a number of learning and fixed agents. They are specified in the .yaml files as Agent1 and Agent2. Canonically, we care about the outcome of Agent1.
We provide an agent interface to help develop new agents for pax. These methods are necessary for existing runners and should be functional.
All the learning strategies have their own folder and the fixed agents can be viewed in pax/agents/strategies.py (inspired by Axelrods' Tournament).
Below we list currently supported agents and their respective implementation details
| Agents | Description |
|---|---|
Naive |
A learning agent that updates using REINFORCE [insert paper here] |
PPO |
A learning agent that updates using PPO [insert paper here] |
PPO_memory |
A learning agent that updates using PPO and a memory state [insert paper here] |
MFOS |
Model free opponent shaping meta-agent based upon the ICML paper 2021 |
GS |
The Good Shepard meta-agent based upon the paper |
Defect |
A fixed agent that always defects |
Altruistic |
A fixed agent that always cooperates |
TitForTat |
A fixed agent that cooperates on the first move and then reciprocates action of the opponent from the previous turn |
*Note: MFOS and GS are meta-agents, so will only work with the meta environment