Learning in Network Games

We report the findings of an experiment designed to study how people learn and make decisions in network games. Network games offer new opportunities to identify learning rules, since on networks (compared to e.g. random matching) more rules differ in terms of their information requirements. Our experimental design enables us to observe both which actions participants choose and which information they consult before making their choices. We use this information to estimate learning types using maximum likelihood methods. There is substantial heterogeneity in learning types. However, the vast majority of our participants' decisions are best characterized by reinforcement learning or (myopic) best-response learning. The distribution of learning types seems fairly stable across contexts. Neither network topology nor the position of a player in the network seem to substantially affect the estimated distribution of learning types.

Learning Multirobot Hose Transportation and Deployment by Distributed Round-Robin Q-Learning

Multi-Agent Reinforcement Learning (MARL) algorithms face two main difficulties: the curse of dimensionality, and environment non-stationarity due to the independent learning processes carried out by the agents concurrently. In this paper we formalize and prove the convergence of a Distributed Round Robin Q-learning (D-RR-QL) algorithm for cooperative systems. The computational complexity of this algorithm increases linearly with the number of agents. Moreover, it eliminates environment non sta tionarity by carrying a round-robin scheduling of the action selection and execution. That this learning scheme allows the implementation of Modular State-Action Vetoes (MSAV) in cooperative multi-agent systems, which speeds up learning convergence in over-constrained systems by vetoing state-action pairs which lead to undesired termination states (UTS) in the relevant state-action subspace. Each agent's local state-action value function learning is an independent process, including the MSAV policies. Coordination of locally optimal policies to obtain the global optimal joint policy is achieved by a greedy selection procedure using message passing. We show that D-RR-QL improves over state-of-the-art approaches, such as Distributed Q-Learning, Team Q-Learning and Coordinated Reinforcement Learning in a paradigmatic Linked Multi-Component Robotic System (L-MCRS) control problem: the hose transportation task. L-MCRS are over-constrained systems with many UTS induced by the interaction of the passive linking element and the active mobile robots.

Structure learning in a sensorimotor association task.

Learning is often understood as an organism's gradual acquisition of the association between a given sensory stimulus and the correct motor response. Mathematically, this corresponds to regressing a mapping between the set of observations and the set of actions. Recently, however, it has been shown both in cognitive and motor neuroscience that humans are not only able to learn particular stimulus-response mappings, but are also able to extract abstract structural invariants that facilitate generalization to novel tasks. Here we show how such structure learning can enhance facilitation in a sensorimotor association task performed by human subjects. Using regression and reinforcement learning models we show that the observed facilitation cannot be explained by these basic models of learning stimulus-response associations. We show, however, that the observed data can be explained by a hierarchical Bayesian model that performs structure learning. In line with previous results from cognitive tasks, this suggests that hierarchical Bayesian inference might provide a common framework to explain both the learning of specific stimulus-response associations and the learning of abstract structures that are shared by different task environments.

Effects of the user model on simulation-based learning of dialogue strategies

Over the past decade, a variety of user models have been proposed for user simulation-based reinforcement-learning of dialogue strategies. However, the strategies learned with these models are rarely evaluated in actual user trials and it remains unclear how the choice of user model affects the quality of the learned strategy. In particular, the degree to which strategies learned with a user model generalise to real user populations has not be investigated. This paper presents a series of experiments that qualitatively and quantitatively examine the effect of the user model on the learned strategy. Our results show that the performance and characteristics of the strategy are in fact highly dependent on the user model. Furthermore, a policy trained with a poor user model may appear to perform well when tested with the same model, but fail when tested with a more sophisticated user model. This raises significant doubts about the current practice of learning and evaluating strategies with the same user model. The paper further investigates a new technique for testing and comparing strategies directly on real human-machine dialogues, thereby avoiding any evaluation bias introduced by the user model. © 2005 IEEE.

Altruistic learning.

The origin of altruism remains one of the most enduring puzzles of human behaviour. Indeed, true altruism is often thought either not to exist, or to arise merely as a miscalculation of otherwise selfish behaviour. In this paper, we argue that altruism emerges directly from the way in which distinct human decision-making systems learn about rewards. Using insights provided by neurobiological accounts of human decision-making, we suggest that reinforcement learning in game-theoretic social interactions (habitisation over either individuals or games) and observational learning (either imitative of inference based) lead to altruistic behaviour. This arises not only as a result of computational efficiency in the face of processing complexity, but as a direct consequence of optimal inference in the face of uncertainty. Critically, we argue that the fact that evolutionary pressure acts not over the object of learning ('what' is learned), but over the learning systems themselves ('how' things are learned), enables the evolution of altruism despite the direct threat posed by free-riders.

Opponent appetitive-aversive neural processes underlie predictive learning of pain relief.

Termination of a painful or unpleasant event can be rewarding. However, whether the brain treats relief in a similar way as it treats natural reward is unclear, and the neural processes that underlie its representation as a motivational goal remain poorly understood. We used fMRI (functional magnetic resonance imaging) to investigate how humans learn to generate expectations of pain relief. Using a pavlovian conditioning procedure, we show that subjects experiencing prolonged experimentally induced pain can be conditioned to predict pain relief. This proceeds in a manner consistent with contemporary reward-learning theory (average reward/loss reinforcement learning), reflected by neural activity in the amygdala and midbrain. Furthermore, these reward-like learning signals are mirrored by opposite aversion-like signals in lateral orbitofrontal cortex and anterior cingulate cortex. This dual coding has parallels to 'opponent process' theories in psychology and promotes a formal account of prediction and expectation during pain.

Bayesian Learning of Noisy Markov Decision Processes

We consider the inverse reinforcement learning problem, that is, the problem of learning from, and then predicting or mimicking a controller based on state/action data. We propose a statistical model for such data, derived from the structure of a Markov decision process. Adopting a Bayesian approach to inference, we show how latent variables of the model can be estimated, and how predictions about actions can be made, in a unified framework. A new Markov chain Monte Carlo (MCMC) sampler is devised for simulation from the posterior distribution. This step includes a parameter expansion step, which is shown to be essential for good convergence properties of the MCMC sampler. As an illustration, the method is applied to learning a human controller.

Behavior learning of a face robot based on the characteristics of human instruction

Computer simulation experiments were performed to examine the effectiveness of OR- and comparative-reinforcement learning algorithms. In the simulation, human rewards were given as +1 and -1. Two models of human instruction that determine which reward is to be given in every step of a human instruction were used. Results show that human instruction may have a possibility of including both model-A and model-B characteristics, and it can be expected that the comparative-reinforcement learning algorithm is more effective for learning by human instructions.

Learning with Deictic Representation

Most reinforcement learning methods operate on propositional representations of the world state. Such representations are often intractably large and generalize poorly. Using a deictic representation is believed to be a viable alternative: they promise generalization while allowing the use of existing reinforcement-learning methods. Yet, there are few experiments on learning with deictic representations reported in the literature. In this paper we explore the effectiveness of two forms of deictic representation and a naive propositional representation in a simple blocks-world domain. We find, empirically, that the deictic representations actually worsen performance. We conclude with a discussion of possible causes of these results and strategies for more effective learning in domains with objects.

SOVEREIGN: An Autonomous Neural System for Incrementally Learning Planned Action Sequence to Navigate Towards a Rewarded Goal

How do reactive and planned behaviors interact in real time? How are sequences of such behaviors released at appropriate times during autonomous navigation to realize valued goals? Controllers for both animals and mobile robots, or animats, need reactive mechanisms for exploration, and learned plans to reach goal objects once an environment becomes familiar. The SOVEREIGN (Self-Organizing, Vision, Expectation, Recognition, Emotion, Intelligent, Goaloriented Navigation) animat model embodies these capabilities, and is tested in a 3D virtual reality environment. SOVEREIGN includes several interacting subsystems which model complementary properties of cortical What and Where processing streams and which clarify similarities between mechanisms for navigation and arm movement control. As the animat explores an environment, visual inputs are processed by networks that are sensitive to visual form and motion in the What and Where streams, respectively. Position-invariant and sizeinvariant recognition categories are learned by real-time incremental learning in the What stream. Estimates of target position relative to the animat are computed in the Where stream, and can activate approach movements toward the target. Motion cues from animat locomotion can elicit head-orienting movements to bring a new target into view. Approach and orienting movements are alternately performed during animat navigation. Cumulative estimates of each movement are derived from interacting proprioceptive and visual cues. Movement sequences are stored within a motor working memory. Sequences of visual categories are stored in a sensory working memory. These working memories trigger learning of sensory and motor sequence categories, or plans, which together control planned movements. Predictively effective chunk combinations are selectively enhanced via reinforcement learning when the animat is rewarded. Selected planning chunks effect a gradual transition from variable reactive exploratory movements to efficient goal-oriented planned movement sequences. Volitional signals gate interactions between model subsystems and the release of overt behaviors. The model can control different motor sequences under different motivational states and learns more efficient sequences to rewarded goals as exploration proceeds.

The Hippocampus and Cerebellum in Adaptively Timed Learning, Recognition, and Movement

The concepts of declarative memory and procedural memory have been used to distinguish two basic types of learning. A neural network model suggests how such memory processes work together as recognition learning, reinforcement learning, and sensory-motor learning take place during adaptive behaviors. To coordinate these processes, the hippocampal formation and cerebellum each contain circuits that learn to adaptively time their outputs. Within the model, hippocampal timing helps to maintain attention on motivationally salient goal objects during variable task-related delays, and cerebellar timing controls the release of conditioned responses. This property is part of the model's description of how cognitive-emotional interactions focus attention on motivationally valued cues, and how this process breaks down due to hippocampal ablation. The model suggests that the hippocampal mechanisms that help to rapidly draw attention to salient cues could prematurely release motor commands were not the release of these commands adaptively timed by the cerebellum. The model hippocampal system modulates cortical recognition learning without actually encoding the representational information that the cortex encodes. These properties avoid the difficulties faced by several models that propose a direct hippocampal role in recognition learning. Learning within the model hippocampal system controls adaptive timing and spatial orientation. Model properties hereby clarify how hippocampal ablations cause amnesic symptoms and difficulties with tasks which combine task delays, novelty detection, and attention towards goal objects amid distractions. When these model recognition, reinforcement, sensory-motor, and timing processes work together, they suggest how the brain can accomplish conditioning of multiple sensory events to delayed rewards, as during serial compound conditioning.

Simulating Effects of Learning and Lesions with a Model of Intrinsic and Synaptically Gated Responses of Striatal Cholinergic Interneurons

The giant cholinergic interneurons of the striatum are tonically active neurons (TANs) that respond with characteristic pauses to novel events and to appetitive and aversive conditioned stimuli. Fluctuations in acetylcholine release by TANs modulate performance- and learning-related dynamics in the striatum. Whereas tonic activity emerges from intrinsic properties of these neurons, glutamatergic inputs from thalamic centromedian-parafascicular nuclei, and dopaminergic inputs from midbrain, are required for the generation of pause responses. No prior computational models encompass both intrinsic and synaptically-gated dynamics. We present a mathematical model that robustly accounts for behavior-related electrophysiological properties of TANs in terms of their intrinsic physiological properties and known afferents. In the model, balanced intrinsic hyperpolarizing and depolarizing currents engender tonic firing, and glutamatergic inputs from thalamus (and cortex) both directly excite and indirectly inhibit TANs. If the latter inhibition, presumably mediated by GABAergic interneurons, exceeds a threshold, its effect is amplified by a KIR current to generate a prolonged pause. In the model, the intrinsic mechanisms and external inputs are both modulated by learning-dependent dopamine (DA) signals and our simulations revealed that many learning-dependent behaviors of TANs are explicable without recourse to learning-dependent changes in synapses onto TANs. The "teaching signal" that modulates reinforcement learning at cortico-striatal synapses may be a sequence composed of an adaptively scaled DA burst, a brief ACh burst, and a scaled ACh pause. Such an interpretation is consistent with recent data on cholinergic control of LTD of cortical synapses onto striatal spiny projection neurons.

A heterosynaptic learning rule for neural networks

In this article we intoduce a novel stochastic Hebb-like learning rule for neural networks that is neurobiologically motivated. This learning rule combines features of unsupervised (Hebbian) and supervised (reinforcement) learning and is stochastic with respect to the selection of the time points when a synapse is modified. Moreover, the learning rule does not only affect the synapse between pre- and postsynaptic neuron, which is called homosynaptic plasticity, but effects also further remote synapses of the pre-and postsynaptic neuron. This more complex form of synaptic plasticity has recently come under investigations in neurobiology and is called heterosynaptic plasticity. We demonstrate that this learning rule is useful in training neural networks by learning parity functions including the exclusive-or (XOR) mapping in a multilayer feed-forward network. We find, that our stochastic learning rule works well, even in the presence of noise. Importantly, the mean leaxning time increases with the number of patterns to be learned polynomially, indicating efficient learning.

Intelligent decision making in electricity markets: simulated annealing Q-Learning

Electricity markets are complex environments, involving a large number of different entities, playing in a dynamic scene to obtain the best advantages and profits. MASCEM is a multi-agent electricity market simulator to model market players and simulate their operation in the market. Market players are entities with specific characteristics and objectives, making their decisions and interacting with other players. MASCEM is integrated with ALBidS, a system that provides several dynamic strategies for agents’ behavior. This paper presents a method that aims at enhancing ALBidS competence in endowing market players with adequate strategic bidding capabilities, allowing them to obtain the higher possible gains out of the market. This method uses a reinforcement learning algorithm to learn from experience how to choose the best from a set of possible actions. These actions are defined accordingly to the most probable points of bidding success. With the purpose of accelerating the convergence process, a simulated annealing based algorithm is included.

Adaptive learning in agents behaviour: a framework for electricity markets simulation

Electricity markets are complex environments with very particular characteristics. A critical issue regarding these specific characteristics concerns the constant changes they are subject to. This is a result of the electricity markets’ restructuring, which was performed so that the competitiveness could be increased, but it also had exponential implications in the increase of the complexity and unpredictability in those markets scope. The constant growth in markets unpredictability resulted in an amplified need for market intervenient entities in foreseeing market behaviour. The need for understanding the market mechanisms and how the involved players’ interaction affects the outcomes of the markets, contributed to the growth of usage of simulation tools. Multi-agent based software is particularly well fitted to analyze dynamic and adaptive systems with complex interactions among its constituents, such as electricity markets. This dissertation presents ALBidS – Adaptive Learning strategic Bidding System, a multiagent system created to provide decision support to market negotiating players. This system is integrated with the MASCEM electricity market simulator, so that its advantage in supporting a market player can be tested using cases based on real markets’ data. ALBidS considers several different methodologies based on very distinct approaches, to provide alternative suggestions of which are the best actions for the supported player to perform. The approach chosen as the players’ actual action is selected by the employment of reinforcement learning algorithms, which for each different situation, simulation circumstances and context, decides which proposed action is the one with higher possibility of achieving the most success. Some of the considered approaches are supported by a mechanism that creates profiles of competitor players. These profiles are built accordingly to their observed past actions and reactions when faced with specific situations, such as success and failure. The system’s context awareness and simulation circumstances analysis, both in terms of results performance and execution time adaptation, are complementary mechanisms, which endow ALBidS with further adaptation and learning capabilities.