An echo state model of non-markovian reinforcement learning

There exists a growing need for intelligent, autonomous control strategies that operate in real-world domains. Theoretically the state-action space must exhibit the Markov property in order for reinforcement learning to be applicable. Empirical evidence, however, suggests that reinforcement learning also applies to domains where the state-action space is approximately Markovian, a requirement for the overwhelming majority of real-world domains. These domains, termed non-Markovian reinforcement learning domains, raise a unique set of practical challenges. The reconstruction dimension required to approximate a Markovian state-space is unknown a priori and can potentially be large. Further, spatial complexity of local function approximation of the reinforcement learning domain grows exponentially with the reconstruction dimension. Parameterized dynamic systems alleviate both embedding length and state-space dimensionality concerns by reconstructing an approximate Markovian state-space via a compact, recurrent representation. Yet this representation extracts a cost; modeling reinforcement learning domains via adaptive, parameterized dynamic systems is characterized by instability, slow-convergence, and high computational or spatial training complexity. The objectives of this research are to demonstrate a stable, convergent, accurate, and scalable model of non-Markovian reinforcement learning domains. These objectives are fulfilled via fixed point analysis of the dynamics underlying the reinforcement learning domain and the Echo State Network, a class of parameterized dynamic system [30]. Understanding models of non-Markovian reinforcement learning domains requires understanding the interactions between learning domains and their models. Fixed point analysis of the Mountain Car Problem reinforcement learning domain, for both local and nonlocal function approximations, suggests a close relationship between the locality of the approximation and the number and severity of bifurcations of the fixed point structure. This research suggests the likely cause of this relationship: reinforcement learning domains exist within a dynamic feature space in which trajectories are analogous to states. The fixed point structure maps dynamic space onto state-space. This explanation suggests two testable hypotheses. Reinforcement learning is sensitive to state-space locality because states cluster as trajectories in time rather than space. Second, models using trajectory-based features should exhibit good modeling performance and few changes in fixed point structure. Analysis of performance of lookup table, feedforward neural network, and Echo State Network (ESN) on the Mountain Car Problem reinforcement learning domain confirm these hypotheses. The ESN is a large, sparse, randomly-generated, unadapted recurrent neural network, which adapts a linear projection of the target domain onto the hidden layer. ESN modeling results on reinforcement learning domains show it achieves performance comparable to lookup table and neural network architectures on the Mountain Car Problem with minimal changes to fixed point structure. Also, the ESN achieves lookup table caliber performance when modeling Acrobot, a four-dimensional control problem, but is less successful modeling the lower dimensional Modified Mountain Car Problem. These performance discrepancies are attributed to the ESN's excellent ability to represent complex short term dynamics, and its inability to consolidate long temporal dependencies into a static memory. Without memory consolidation, reinforcement learning domains exhibiting attractors with multiple dynamic scales are unlikely to be well-modeled via ESN. To mediate this problem, a simple ESN memory consolidation method is presented and tested for stationary dynamic systems. These results indicate the potential to improve modeling performance in reinforcement learning domains via memory consolidation.