Stafford, Ronnie James; (2020) Improving the Practicality of Model-Based Reinforcement Learning: An Investigation into Scaling up Model-Based Methods in Online Settings. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Stafford_10089835_thesis.pdf Download (15MB) | Preview

Abstract

This thesis is a response to the current scarcity of practical model-based control algorithms in the reinforcement learning (RL) framework. As of yet there is no consensus on how best to integrate imperfect transition models into RL whilst mitigating policy improvement instabilities in online settings. Current state-of-the-art policy learning algorithms that surpass human performance often rely on model-free approaches that enjoy unmitigated sampling of transition data. Model-based RL (MBRL) instead attempts to distil experience into transition models that allow agents to plan new policies without needing to return to the environment and sample more data. The initial focus of this investigation is on kernel conditional mean embeddings (CMEs) (Song et al., 2009) deployed in an approximate policy iteration (API) algorithm (Grünewälder et al., 2012a). This existing MBRL algorithm boasts theoretically stable policy updates in continuous state and discrete action spaces. The Bellman operator’s value function and (transition) conditional expectation are modelled and embedded respectively as functions in a reproducing kernel Hilbert space (RKHS). The resulting finite-induced approximate pseudo-MDP (Yao et al., 2014a) can be solved exactly in a dynamic programming algorithm with policy improvement suboptimality guarantees. However model construction and policy planning scale cubically and quadratically respectively with the training set size, rendering the CME impractical for sampleabundant tasks in online settings. Three variants of CME API are investigated to strike a balance between stable policy updates and reduced computational complexity. The first variant models the value function and state-action representation explicitly in a parametric CME (PCME) algorithm with favourable computational complexity. However a soft conservative policy update technique is developed to mitigate policy learning oscillations in the planning process. The second variant returns to the non-parametric embedding and contributes (along with external work) to the compressed CME (CCME); a sparse and computationally more favourable CME. The final variant is a fully end-to-end differentiable embedding trained with stochastic gradient updates. The value function remains modelled in an RKHS such that backprop is driven by a non-parametric RKHS loss function. Actively compressed CME (ACCME) satisfies the pseudo-MDP contraction constraint using a sparse softmax activation function. The size of the pseudo-MDP (i.e. the size of the embedding’s last layer) is controlled by sparsifying the last layer weight matrix by extending the truncated gradient method (Langford et al., 2009) with group lasso updates in a novel ‘use it or lose it’ neuron pruning mechanism. Surprisingly this technique does not require extensive fine-tuning between control tasks.

Download activity - last month

Export as XMLJSONCSV

Download activity - last 12 months

Export as CSVJSONXML

Downloads by country - last 12 months

Export as CSVJSONXML

Archive Staff Only

View Item