bandofstraycats / naq

Implementation of Approximate Smooth Kernel Value Iteration

Implementation of Approximate Smooth Kernel Value Iteration

Repository

• kernel_vi.py - Approximate Smooth Kernel Value Iteration
• kernel.py - Kernel definition
• gridworld_mdp.py - GridWorld domain
• plot.py - Plot performance metrics of one run
• plot.R - Plot performance metrics across runs
• kernel_vi.sh - Generates multiple runs across random seeds

Installation

Requirements: Python, numpy, matplotlib

Usage

Description of parameters is provided in the help message python kernel_vi.py --help

Example on a stochastic cliff walking problem

Value Iteration

• Value Iteration python kernel_vi.py --plan plans/plan0.txt --random-slide 0.15 --opt-v plans/opt_v0_rew_5_rs_0.15.txt --max-iter 20 --plot metrics.png
• Approximate Value Iteration with sampled Bellman operator at 10 states python kernel_vi.py --plan plans/plan0.txt --random-slide 0.15 --opt-v plans/opt_v0_rew_5_rs_0.15.txt --max-iter 100 --s 10 --log-steps 10 --plot metrics.png

Kernel Value Iteration

• Kernel Value Iteration with Neural Tangent Kernel python kernel_vi.py --plan plans/plan0.txt --random-slide 0.15 --opt-v plans/opt_v0_rew_5_rs_0.15.txt --max-iter 20 --kernel --kernel-type ntk --plot metrics.png
• Approximate Kernel Value Iteration with Neural Tangent Kernel and sampled Bellman operator at one state python kernel_vi.py --plan plans/plan0.txt --random-slide 0.15 --opt-v plans/opt_v0_rew_5_rs_0.15.txt --max-iter 20 --s 1 --kernel --kernel-type ntk --plot metrics.png

Aggregate performance metrics

Generates NUM_RUNS runs of Approximate Smooth Kernel Value Iteration across random seeds. Saves performance metrics across iterations into 'export' directory for each seed.

bash kernel_vi.sh 0 NUM_RUNS

Plot performance metrics across runs

Requirements: R-project, install.packages(c('ggplot2', 'reshape2', 'dplyr'))

Reads 'export' directory from previous step and generates plot.pdf in the current directory Rscript plot.R

References

[1] Smirnova, Elena. On Convergence of Neural asynchronous Q-iteration. EWRL, 2022.