Overview

Environment is a concept in Sotopia to orchestrating the interactions between agents. Environments provide a social context for the agents, passing messages between agents and evaluating the quality of social interactions. Environments can be customized to test different scenarios, different turn-ordering, and different ways to evaluate the social interactions.

ParallelSotopiaEnv

The ParallelSotopiaEnv is a general environment that runs multiple agents in parallel. The "parallel" in the name refers to the fact that the agents could run in parallel. However, there are three kinds of ordering supported by the ParallelSotopiaEnv:

1. Random ordering: in each turn, a randomly chosen agent can act.
2. Round-robin ordering: in turn i, agent i % n acts, where n is the number of agents.
3. Simultaneous ordering: All agents take actions at the same time.

Except for 3, in each turn one of agent's aact method is called, while other agents' aact methods are not called.

Evaluators

There are two kinds of evaluators in the ParallelSotopiaEnv:

1. Response evaluators: These evaluators are called after each agent's action.
2. Terminal evaluators: These evaluators are called after the environment reaches a terminal state.

The first kind of evaluator determines whether the interaction should terminate, and the second kind of evaluator determines the final reward for each agent.

One example of the first kind of evaluator is RuleBasedTerminatedEvaluator, which test if the interaction has staled for too long or if a certain number of turns has passed. And one example of the second kind of evaluator is ReachGoalLLMEvaluator[T_eval_dim], which evaluates the agents based on a specification of dimensions passed as the type variable T_eval_dim.

T_eval_dim is a type variable that specifies the dimensions of the evaluation, which should subclass the pydantic.BaseModel class. And example of it is SotopiaDimensions, which is a class that contains the dimensions of the evaluation used in the original Sotopia paper.

Usage of ParallelSotopiaEnv

To use the ParallelSotopiaEnv, you need to create a list of agents and a list of evaluators, and pass them to the constructor of the ParallelSotopiaEnv. Then you can call the astep method to pass the actions of the agents to the environment. As of now, the simplest way to use the ParallelSotopiaEnv is to use the run_async_server function in the sotopia.server module. Check out the minimalist_demo example:

asyncio.run(
    run_async_server(
        model_dict={
            "env": "gpt-4",
            "agent1": "gpt-3.5-turbo",
            "agent2": "gpt-3.5-turbo",
        },
        sampler=UniformSampler(),
    )
)