An OpenAI gym environment to evaluate the ability of large language models (LLMs; eg. GPT-4, Claude) in long-horizon reasoning and task planning in dynamic multi-agent settings based on gym-cooking [1].
https://www.youtube.com/watch?v=4LmcpkS53Wg
There is a new area of AI research where foundation models such as LLMs are used for decision making in complex environments that involve long-horizon reasoning, control, and planning [2]. For instance, Text2Motion [3] enables robots to solve sequential manipulation tasks by using LLMs. Also OpenAI's GPT-4 performs well in theory-of-mind (ToM) tasks [6], which require understanding other agents' beliefs, goals, and mental states.
OvercookedGPT is an interactive 2D game environment where OpenAI's GPT-4/3.5-Turbo as well as Anthropic's Claude generate intertemporal and sequential tasks in a centralized fashion to control multiple agents to achieve a goal in a simulation (i.e., to cook food at a kitchen). It is based on gym-cooking [1] and was also inspired by overcooked_ai [4] (which is used in [5]). The purpose of this simulator is to evaluate the ability of the LLMs in long-horizon reasoning and task planning in dynamic multi-agent environments. To this end, in-context learning (i.e., few-shot learning with prompt engineering methods of CoT and PAL [7]) is used to guide the LLMs to generate a task queue in Python that is executed by the simulator on the fly. As shown in [8], the reasoning performance of LLMs improves with an increased degree of input prompt complexity, where complex prompts achieve better performance than simple ones.
OvercookedGPT uses ICL (in-context learning) to guide LLMs to generate a task queue to control multiple agents, but why/how can LLMs (whose parameters including the W_Q, W_K and W_V matrices in the Transformer attention are "frozen") perform an unseen, non-pretrained language task by merely observing a few demonstrations? [9] shows that LLMs have the ability to override semantic priors when presented with demonstration examples (input-label pairs) despite the stronger semantic priors that the models may hold (which ability is an emergent phenomena unlocked by model scaling). Moreover, [10] suggests that LLMs perform gradient descent in ICL through forward computation (whereas explicit finetuning computes gradients by back-propagation) - ICL can therefore be considered as implicit finetuning.
python3 -m pip install -U pygame --user
git clone https://github.com/BladeTransformerLLC/OvercookedGPT.git
cd OvercookedGPT
pip3 install -r requirements.txt
Set the OPENAI_API_KEY environment variable (or put the key string in utils/chatgpt/openai.json).
Set the ANTHROPIC_API_KEY environment variable (or put the key string in utils/claude/anthropic.json).
Start a single-agent simulation using GPT-3.5-Turbo by default (specify a model in utils/chatgpt/openai.json):
python3 main.py --num-agents 1 --level partial-divider_salad --gpt
(or replace --gpt with --claude for using Anthropic's Claude). Enter a task eg. "Make a tomato and lettuce salad and deliver it."
Start a multi-agent simulation:
python3 main.py --num-agents 2 --level partial-divider_salad --gpt
Mannually control agents with arrow keys (switch between agents by pressing 1 or 2):
python3 main.py --num-agents 2 --level partial-divider_salad --gpt --manual
- Allow simultaneous/parallel subtask execution by multiple agents within the same timestep
- Prevent agents from moving through other agents (make them avoid/wait others)
- Evaluate with 3 or more agents in large levels
- Wu et al., "Too many cooks: Bayesian inference for coordinating multi-agent collaboration," 2020.
- Yang et al., "Foundation Models for Decision Making: Problems, Methods, and Opportunities," 2023.
- Lin et al., "Text2Motion: From Natural Language Instructions to Feasible Plans," 2023.
- Carroll et al., "On the Utility of Learning about Humans for Human-AI Coordination," 2020.
- Hong et al., "Learning to Influence Human Behavior with Offline Reinforcement Learning," 2023.
- Moghaddam & Honey, "Boosting Theory-of-Mind Performance in Large Language Models via Prompting," 2023.
- Gao et al., "PAL: Program-aided Language Models," 2022.
- Fu et al., "Complexity-Based Prompting for Multi-Step Reasoning," 2022.
- Wei et al., "Larger language models do in-context learning differently," 2023.
- Dai et al., "Why Can GPT Learn In-Context? Language Models Secretly Perform Gradient Descent as Meta-Optimizers," 2022.