Trust-Region Twisted Policy Improvement
This work addresses the problem of inefficient online planning in reinforcement learning for researchers and practitioners, offering incremental improvements over existing MCTS and SMC methods.
The paper tackled the challenge of scaling Monte-Carlo tree search (MCTS) for parallel compute in deep reinforcement learning by developing Trust-Region Twisted SMC (TRT-SMC), a tailored sequential Monte-Carlo planner that improved runtime and sample-efficiency over baseline methods in discrete and continuous domains.
Monte-Carlo tree search (MCTS) has driven many recent breakthroughs in deep reinforcement learning (RL). However, scaling MCTS to parallel compute has proven challenging in practice which has motivated alternative planners like sequential Monte-Carlo (SMC). Many of these SMC methods adopt particle filters for smoothing through a reformulation of RL as a policy inference problem. Yet, persisting design choices of these particle filters often conflict with the aim of online planning in RL, which is to obtain a policy improvement at the start of planning. Drawing inspiration from MCTS, we tailor SMC planners specifically for RL by improving data generation within the planner through constrained action sampling and explicit terminal state handling, as well as improving policy and value target estimation. This leads to our Trust-Region Twisted SMC (TRT-SMC), which shows improved runtime and sample-efficiency over baseline MCTS and SMC methods in both discrete and continuous domains.