import glob import os import gymnasium as gym import numpy as np from gymnasium.wrappers import RecordVideo from moviepy.video.compositing.concatenate import concatenate_videoclips from moviepy.video.io.VideoFileClip import VideoFileClip from sympy import latex from interpretable import InterpretablePolicyExtractor from utils import generate_dataset_from_expert, rollouts import matplotlib.pyplot as plt import torch import gradio as gr intro = """ # Making RL Policy Interpretable with Kolmogorov-Arnold Network 🧠 ➙ 🔢 Waris Radji1, Corentin Léger2, Hector Kohler1 1[Inria, team Scool](https://team.inria.fr/scool/) 2[Inria, team Flowers](https://flowers.inria.fr/) In this demo, we showcase a method to make a trained Reinforcement Learning (RL) policy interpretable using the Kolmogorov-Arnold Network (KAN). The process involves transferring the knowledge from a pre-trained RL policy to a KAN. We achieve this by training the KAN to map actions from observations obtained from trajectories of the pre-trained policy. ## Procedure - Train the KAN using observations from trajectories generated by a pre-trained RL policy, the KAN learns to map observations to corresponding actions. - Apply symbolic regression algorithms to the KAN's learned mapping. - Extract an interpretable policy expressed in symbolic form. For more information about KAN you can read the [paper](https://arxiv.org/abs/2404.19756), and check the [PyTorch official information](https://github.com/KindXiaoming/pykan). To follow the progress of KAN in RL you can check the repo [kanrl](https://github.com/riiswa/kanrl). [![riiswa/kanrl - GitHub](https://gh-card.dev/repos/riiswa/kanrl.svg)](https://github.com/riiswa/kanrl) *Please be patient, as the process may take a few minutes to run, especially in environments with large state/action spaces or with a complex KAN architecture. Some Mujoco environments are available if you run this app locally. For optimal performance, default parameters may not suffice. Feel free to experiment with different settings to achieve desired results.* **If the space is to busy, don't hesitate to do your experiments [locally](https://github.com/riiswa/kanrl/tree/main/interpretable).** """ envs = ["CartPole-v1", "MountainCar-v0", "Acrobot-v1", "Pendulum-v1", "MountainCarContinuous-v0", "LunarLander-v2", "BipedalWalker-v3"] #"Swimmer-v3", "Hopper-v3", "HalfCheetah-v3", "Walker2d-v3"] if __name__ == "__main__": torch.set_default_dtype(torch.float32) def load_video_and_dataset(_env_name): env_name = _env_name agent = "ppo" if env_name == "Swimmer-v3" or env_name == "Walker2d-v3": agent = "trpo" dataset_path, video_path = generate_dataset_from_expert(agent, _env_name, 25, 3) return video_path, gr.Button("Compute the symbolic policy!", interactive=True), { "dataset_path": dataset_path, "ipe": None, "env_name": env_name } def parse_integer_list(input_str): if not input_str or input_str.isspace(): return None elements = input_str.split(',') try: int_list = tuple([int(elem.strip()) for elem in elements]) return int_list except ValueError: return False def extract_interpretable_policy(kan_widths, epochs, state): widths = parse_integer_list(kan_widths) if kan_widths is False: gr.Warning(f"Please enter widths {kan_widths} in the right format... The current run is executed with no hidden layer.") widths = None state["ipe"] = InterpretablePolicyExtractor(state["env_name"], widths) state["ipe"].train_from_dataset(state["dataset_path"], steps=epochs) state["ipe"].policy.prune() state["ipe"].policy.plot(mask=True, scale=5) fig = plt.gcf() fig.canvas.draw() kan_architecture = np.array(fig.canvas.renderer.buffer_rgba()) plt.close() return kan_architecture, state, fig def symbolic_policy(state): lib = ['x', 'x^2', 'x^3', 'x^4', 'exp', 'log', 'sqrt', 'tanh', 'sin', 'abs'] state["ipe"].policy.auto_symbolic(lib=lib) env = gym.make(state["env_name"], render_mode="rgb_array") env = RecordVideo(env, video_folder="videos", episode_trigger=lambda x: True, name_prefix=f"""kan-{state["env_name"]}""") rollouts(env, state["ipe"].forward, 2) video_path = os.path.join("videos", f"""kan-{state["env_name"]}.mp4""") video_files = glob.glob(os.path.join("videos", f"""kan-{state["env_name"]}-episode*.mp4""")) clips = [VideoFileClip(file) for file in video_files] final_clip = concatenate_videoclips(clips) final_clip.write_videofile(video_path, codec="libx264", fps=24) symbolic_formula = f"### The symbolic formula of the policy is:" formulas = state["ipe"].policy.symbolic_formula()[0] for i, formula in enumerate(formulas): symbolic_formula += "\n$$ a_" + str(i) + "=" + latex(formula) + "$$" if state["ipe"]._action_is_discrete: symbolic_formula += "\n" + r"$$ a = \underset{i}{\mathrm{argmax}} \ a_i.$$" return video_path, symbolic_formula, gr.Button("Compute the symbolic policy!", interactive=True) css = """ #formula {overflow-x: auto!important}; """ with gr.Blocks(theme='gradio/monochrome', css=css) as app: state = gr.State({ "dataset_path": None, "ipe": None, "env_name": None }) gr.Markdown(intro) with gr.Row(): with gr.Column(): gr.Markdown("### Pretrained policy loading (PPO or TRPO from [rl-baselines3-zoo](https://github.com/DLR-RM/rl-baselines3-zoo))") choice = gr.Dropdown(envs, label="Environment name") expert_video = gr.Video(label="Expert policy video", interactive=False, autoplay=True) kan_widths = gr.Textbox(value="2", label="Widths of the hidden layers of the KAN, separated by commas (e.g. `3,3`). Leave empty if there are no hidden layers.") epochs = gr.Number(value=20, label="KAN training Steps.", minimum=1, maximum=100) button = gr.Button("Compute the symbolic policy!", interactive=False) with gr.Column(): gr.Markdown("### Symbolic policy extraction") kan_architecture = gr.Image(interactive=False, label="KAN architecture") sym_video = gr.Video(label="Symbolic policy video", interactive=False, autoplay=True) sym_formula = gr.Markdown(elem_id="formula") choice.input(load_video_and_dataset, inputs=[choice], outputs=[expert_video, button, state], concurrency_limit=8) button.click(lambda: gr.Button("Computation in progress...", interactive=False), None, [button]).then(extract_interpretable_policy, inputs=[kan_widths, epochs, state], outputs=[kan_architecture, state]).then( symbolic_policy, inputs=[state], outputs=[sym_video, sym_formula, button] ) app.queue(default_concurrency_limit=3, max_size=10) app.launch(show_error=True)